[{"content":"Silicon, not retail shelves, drove this week\u0026rsquo;s news (Sept 19–25). We could not confirm a single new Korea-market launch this week, but Qualcomm made plenty of noise abroad, unveiling two next-generation flagship chips on a 2nm process at its Snapdragon Summit. On the software side, Alibaba\u0026rsquo;s multimodal Qwen3.8-27B topped Hugging Face\u0026rsquo;s weekly trending chart.\nKorea No confirmed Korea market launch this week. We checked manufacturer press releases and major Korean tech outlets for anything dated inside the Sept 19–25 window and found nothing that qualifies as a new product launch — only older coverage of products already reported in earlier briefs, plus speculative reporting about a possible Xiaomi 18 Korea release later in the year that has not been confirmed as a launch and is not counted here.\nGlobal Snapdragon 8 Elite Gen 6 / Extreme Gen 6 — Qualcomm unveiled two next-generation flagship mobile chips on September 22 (local time) at Snapdragon Summit 2026 in Maui, Hawaii.\nProcess node: 2nm. Snapdragon 8 Elite Gen 6 pairs a 5GHz dual-core Oryon Prime cluster with six 4GHz Performance cores. Qualcomm says it delivers a 10% CPU performance gain and a 37% power-efficiency improvement over the Elite Gen 5. The higher Extreme Gen 6 (Pro trim) uses a 2+3+3 CPU cluster that peaks at 5.11GHz, versus 4.4GHz on the standard chip. Its Adreno 850 GPU carries 18MB of graphics memory, up from the prior Adreno A845\u0026rsquo;s 12MB, for a claimed 44% GPU performance gain and 40% efficiency improvement. Memory and storage: LPDDR6 and UFS 5.0 support. This is the first time Qualcomm has split its flagship line into two tiers — Elite and Extreme — which it says gives OEMs more room to differentiate between mainstream flagships and ultra-premium models, rather than shipping a single silicon option across an entire flagship range. Phones built on these chips are expected to roll out starting with 2027 flagships. Whether Korea-market lineups (Galaxy and others) will adopt them, and when, is TBC — Samsung has not confirmed which Snapdragon variant its next Galaxy S generation will use. Nothing here is consumer-facing yet: this is a component announcement, not a device launch. The gap between a chip unveiling in September and phones actually reaching shelves has typically run several months to a year in past Snapdragon cycles, so treat \u0026ldquo;2027 flagships\u0026rdquo; as the earliest realistic window rather than a firm date. Sources: 9to5Google, Droid Life, Deal N Tech Software Qwen3.8-27B — a multimodal model from Alibaba\u0026rsquo;s Qwen family that led Hugging Face\u0026rsquo;s weekly trending list as of September 23, with 16,079 weekly likes and 7.08 million downloads.\nQwen-family models effectively dominate this week\u0026rsquo;s trending list, reportedly accounting for close to half of the top-ranked models across text, image-text, and video categories combined — a sign that Alibaba\u0026rsquo;s open-weight ecosystem, rather than any single US lab, is currently setting the pace on Hugging Face\u0026rsquo;s download and adoption metrics. License, exact parameter architecture, and the existence of an official quantized release are all TBC. We\u0026rsquo;re intentionally not guessing at those details until we can confirm them directly from the model card. At 27B parameters, this model sits above the on-device threshold this brief prioritizes (roughly 8B or below, or a confirmed quantized variant). Without a quantized release, it isn\u0026rsquo;t practical to run directly on a typical laptop or edge device — it\u0026rsquo;s a server/cloud-scale model for now, even though its download numbers suggest broad developer interest. Worth watching alongside it: Hugging Face\u0026rsquo;s broader trending list this week also skewed toward video-generation models and extreme low-bit quantization, both signs that the on-device and edge-AI push covered in past DigitalBrain dev posts is continuing even where this specific model doesn\u0026rsquo;t fit that mold yet. Sources: Hugging Face trending, agents-radar tracking issue (2026-09-23) This Week at a Glance Section Item Key number Korea — No confirmed launch Global Snapdragon 8 Elite / Extreme Gen 6 2nm, peak 5.11GHz Software Qwen3.8-27B 16,079 likes / 7.08M downloads This was a week defined by a single chip announcement. Korea\u0026rsquo;s retail channel has nothing new to show yet, and the phones that actually carry this silicon won\u0026rsquo;t ship until 2027 flagships arrive. In the meantime, the real, faster-moving story is on the software side, where large-model competition keeps circling back to the Qwen family.\nThis Week on DigitalBrain Apple Watch Ultra 4 Costs More, Runs Out Sooner vivo X500 Pro Wins on Battery, Not on Availability For a Camera Add-On, Skip the $299 VENTUNO Q Cloudflare\u0026rsquo;s Skill Says: Unverified Isn\u0026rsquo;t a Vulnerability ","permalink":"https://dibrain.dev/en/2026/09/25/brief-snapdragon-gen6-qwen/","summary":"No confirmed Korea product launches this week. Globally, Qualcomm unveiled the 2nm Snapdragon 8 Elite Gen 6 and Extreme Gen 6 at its Snapdragon Summit, and on the software side Alibaba\u0026rsquo;s Qwen3.8-27B topped Hugging Face\u0026rsquo;s weekly trending chart with over 16,000 likes.","title":"Qualcomm Ships a New Chip, Korea Stays Quiet"},{"content":"This week\u0026rsquo;s GitHub trending list is heavy on tooling built for AI coding agents — skills, harnesses, and middleware meant to make agents like Claude Code, Codex, or Cursor faster and more reliable. Two repos stood out for taking opposite approaches to that same problem. One makes an agent audit code but refuses to count a finding until it\u0026rsquo;s independently verified by a second agent. The other tears the entire agent runtime apart into swappable plugins so developers can rebuild it however they want. Read together, they say something about where \u0026ldquo;agent tooling\u0026rdquo; is headed this year: less about the model itself, more about the scaffolding around it.\nCloudflare security-audit-skill cloudflare/security-audit-skill is MIT-licensed, Node.js-based, and sits at roughly 20.8k stars. It ships without GitHub Release tags — updates land straight on the main branch.\nThe skill turns a coding agent (Claude Code and similar) into a six-phase security auditor:\nRecon — maps architecture, trust boundaries, and input surfaces Coverage-led hunting — isolated agents attack the codebase from different angles (injection, access control, business logic, and more) Candidate validation — a separate agent tries to disprove each finding Structured output — produces a machine-readable findings JSON Independent record verification — yet another agent re-checks the underlying evidence Target-neutral reporting — the final report is built only from findings that survived verification Install with one line:\nnpx skills add https://github.com/cloudflare/security-audit-skill \\ --skill security-audit --global Once installed, it triggers automatically on prompts like \u0026ldquo;security audit this codebase\u0026rdquo; or \u0026ldquo;find security vulnerabilities in ./src\u0026rdquo;. The interesting part isn\u0026rsquo;t step 6, it\u0026rsquo;s steps 3 through 5. A finding from the first pass doesn\u0026rsquo;t make the report just because an agent flagged it — a different agent has to fail to disprove it, and the underlying evidence gets re-checked before it counts. That\u0026rsquo;s a structural answer to the well-known problem of AI-generated security reports padded with false positives: instead of trusting one pass, the pipeline builds disbelief into the process by design, so a plausible-sounding but wrong finding has three separate chances to get thrown out before it ever reaches a human reviewer.\nBest for: teams already using AI agents to write or review code who want an automated security pass inserted before merge.\nDeepSeek Harness deepseek-ai/deepseek-harness is an MIT-licensed, TypeScript agent runtime. Published on August 13, it passed 100k stars within 48 hours and has since climbed past 200k — one of the fastest-growing repos this year. The latest release is v0.1.7-rc.1; the exact release date is TBC (fetched metadata was inconsistent).\nIts design principle is \u0026ldquo;Everything is a Plugin.\u0026rdquo; Built on Cordis, every piece — model adapters, the tool registry, session logs, even the agent loop itself — is a replaceable plugin. In practice that means you\u0026rsquo;re not locked into whatever model or tool set the maintainers shipped by default: swap the model adapter for a different provider, or drop in your own tool registry, without forking the core.\n# quick start npx @deepseek-ai/dsh web # build from source git clone https://github.com/deepseek-ai/deepseek-harness.git cd deepseek-harness pnpm install pnpm run build pnpm dsh web Running it launches a local web UI at http://127.0.0.1:3080. Per the v0.1.7-rc.1 release notes, this build adds experimental Playwright-based browser automation and Computer Use support, plus plugin management through the web UI.\nThe version string tells its own story: -rc.1 means this isn\u0026rsquo;t a 1.0 release yet. Star count and maturity are two different things here.\nBest for: developers who want to assemble their own agent runtime instead of relying on a closed product like Claude Code or Codex — better suited to studying the architecture and cherry-picking plugins than to dropping straight into production.\nWrap-up Both repos ride the same wave — tooling built for AI agents — but pull in opposite directions. security-audit-skill assumes an agent\u0026rsquo;s own findings can\u0026rsquo;t be trusted and adds verification layers; DeepSeek Harness assumes you want to rebuild the agent itself and hands you every piece as a plugin. The audit skill is ready to slot into an existing workflow today; the harness is still at rc, so start by reading the core loop before betting production on it. Pick based on what you\u0026rsquo;re missing: an extra verification pass in front of code you already ship, or a runtime you can actually take apart.\n","permalink":"https://dibrain.dev/en/2026/09/24/dev-security-audit-skill-deepseek-harness/","summary":"This week\u0026rsquo;s GitHub picks: DeepSeek Harness, a plugin-based agent runtime that crossed 200k stars two weeks after launch, and Cloudflare\u0026rsquo;s security-audit-skill, which adds independent verification before an AI agent\u0026rsquo;s finding counts as a real vulnerability.","title":"Cloudflare's Skill Says: Unverified Isn't a Vulnerability"},{"content":"Arduino announced the UNO Media Carrier on September 9, an add-on board that snaps onto the dual-brain UNO Q and its higher-end sibling, the VENTUNO Q, to add two camera inputs, a display connector, and audio I/O in one shot. At $19.25 (€19.89 VAT included), it costs far less than the UNO Q itself. The most practical detail in this launch isn\u0026rsquo;t the connectors — it\u0026rsquo;s that the exact same carrier fits both boards, from the $39 entry-level UNO Q up to the $299 VENTUNO Q.\nBoth host boards were already on the market before this launch. What\u0026rsquo;s new is that Arduino is now shipping an official, documented way to wire a camera and display to either one, rather than leaving builders to source their own MIPI adapters and breakout cables.\nUNO Media Carrier Spec Value Price $19.25 / €19.89 (VAT incl.) Dimensions 68.58 × 53.34mm Camera input Dual MIPI-CSI (two cameras at once) Display One 22-pin MIPI-DSI connector, supports 5\u0026quot;, 8\u0026quot;, 10.1\u0026quot; Waveshare touch panels Audio Three 3.5mm jacks Operating temperature -10°C to 60°C Power Drawn from the host board\u0026rsquo;s JMEDIA/JMISC connectors (exact voltage/current TBC) Compatible boards UNO Q, VENTUNO Q It\u0026rsquo;s a pass-through accessory with no power input of its own, so buying one means first picking a host: the UNO Q ($39 for 2GB, $59 for a 4GB variant shipping later this year, prices vary by region) or the VENTUNO Q ($299).\nSame Carrier, Two Very Different Hosts Spec UNO Q VENTUNO Q SoC Qualcomm QRB2210 (quad-core Cortex-A53, up to 2GHz) Qualcomm Dragonwing IQ8 Real-time MCU STM32U585 STM32H5 RAM/storage 2GB/16GB or 4GB/32GB TBC Price $39 (2GB) / $59 (4GB, shipping later this year) $299 OS Debian Linux runs on the Qualcomm side TBC Media Carrier support Yes Yes The VENTUNO Q\u0026rsquo;s exact memory configuration and benchmarks weren\u0026rsquo;t disclosed in the sources checked for this piece — that\u0026rsquo;s TBC. What is clear is a $260 price gap between the two boards that use the identical camera/display expansion path.\nWhat This Means at the Embedded Level Power: The carrier has no separate power input; it pulls current from the host\u0026rsquo;s JMEDIA/JMISC pins. How much of the UNO Q\u0026rsquo;s power budget two live cameras plus a display consume isn\u0026rsquo;t published anywhere we found — worth checking the datasheet directly before putting this on a battery-powered robot.\nMemory: The base UNO Q ships with 2GB of RAM. Running two MIPI-CSI camera streams simultaneously, let alone adding a vision model on top, could get tight on 2GB — the 4GB variant is the safer pick for camera-heavy builds.\nThermal: The board\u0026rsquo;s rated -10°C to 60°C range is a component spec, not a measurement of what happens with two cameras and a display running inside a sealed robot chassis. No real-world thermal data has been published for that scenario.\nConnectors: MIPI ribbon cables are short and not very tolerant of flexing. Anything mounted on a moving arm or mobile robot will need its own cable strain relief and routing — this carrier doesn\u0026rsquo;t solve that on its own.\nWho Should Buy Which For hobbyist or educational vision projects, a $39 UNO Q plus a $19.25 carrier covers a camera-and-display setup without spending more. The VENTUNO Q\u0026rsquo;s Dragonwing IQ8 is aimed at heavier on-device AI workloads, so paying $260 more just to plug in a camera doesn\u0026rsquo;t make much sense on its own. If the plan involves real-time video inference or fusing multiple sensor streams, that\u0026rsquo;s where the VENTUNO Q\u0026rsquo;s extra headroom starts to matter.\nAvailability No regional retail or local pricing beyond Arduino\u0026rsquo;s own international store was found for this launch. For now, that means shipping costs and import handling from Arduino\u0026rsquo;s store are part of the real price for buyers outside the US and EU.\nWhy This Matters Beyond Arduino The pattern here — a cheap, official accessory that unlocks camera and display I/O across an entire board lineup rather than just one SKU — is worth watching regardless of what you think of Arduino specifically. It lowers the cost of the first step in any vision or HMI project built on these boards, and it means a maker doesn\u0026rsquo;t have to commit to the priciest SKU in a lineup just to get camera support. Whether Arduino keeps extending that same carrier to future boards, or whether this becomes a one-off for the UNO Q family, isn\u0026rsquo;t something the current announcement answers.\nVerdict The UNO Media Carrier isn\u0026rsquo;t a technical breakthrough so much as a cheap, well-executed way to expose standard MIPI interfaces. But the fact that the same $19.25 board plugs into both a $39 and a $299 host is the useful takeaway: starting a camera project doesn\u0026rsquo;t require buying the flagship. The numbers that actually matter for putting this into real hardware — power draw and thermal behavior under load — still need to come from a closer read of Arduino\u0026rsquo;s datasheets.\n","permalink":"https://dibrain.dev/en/2026/09/23/embedded-arduino-uno-media-carrier/","summary":"Arduino\u0026rsquo;s new $19.25 UNO Media Carrier bolts a camera, display, and audio onto both the UNO Q and the pricier VENTUNO Q. Since the same accessory fits either board, we looked at whether a vision project actually needs the expensive one.","title":"For a Camera Add-On, Skip the $299 VENTUNO Q"},{"content":"vivo announced the X500 series in China on September 21, alongside a larger X500 Pro Max variant we aren\u0026rsquo;t covering here. This piece focuses on the X500 Pro, which leads with a 200MP periscope telephoto camera and MediaTek\u0026rsquo;s new 2nm chip. Pre-orders are already open in China, with retail sales starting September 29. Since vivo doesn\u0026rsquo;t officially sell phones in South Korea, we compared it against a phone Korean buyers can actually walk into a store and pick up: the Samsung Galaxy S25 Ultra, currently the closest thing on the market to an apples-to-apples flagship rival.\nvivo X500 Pro Spec vivo X500 Pro Display 6.37-inch 1.5K OLED, 144Hz Chipset MediaTek Dimensity 9600 Pro (2nm) Memory/Storage 12GB RAM, 256GB / 512GB / 1TB (LPDDR6, UFS 5.0) Battery 6,510mAh Charging 90W wired Main camera 50MP (Sony LYT910) Ultra-wide camera 50MP Telephoto camera 200MP periscope Video 17EV TrueDynamic Log, 10-bit 4:2:2, ACES support Price (China) CNY 6,499 (12GB/256GB) to CNY 8,499 (12GB/1TB) Launch Announced September 21, China sales start September 29 The Dimensity 9600 Pro packs 33 billion transistors on a 2nm process and is the first chip to support the next-generation LPDDR6 RAM and UFS 5.0 storage standards — both meaningfully faster than the LPDDR5X and UFS 4.x generation still standard on most 2026 flagships. Main and ultra-wide cameras both sit at 50MP, but the telephoto lens gets a 200MP sensor in a periscope housing, putting the design weight behind zoom performance rather than wide-angle detail. On video, the phone supports 17-stop dynamic range Log shooting and the ACES color standard, features usually reserved for cinema cameras, which widens how much color grading room you have after the fact rather than baking a fixed look into the footage. The three storage tiers top out at 1TB, aimed squarely at buyers who shoot a lot of high-bitrate Log video and don\u0026rsquo;t want to manage offloading footage mid-trip.\nAgainst the Galaxy S25 Ultra Spec vivo X500 Pro Galaxy S25 Ultra Display 6.37-inch, 144Hz 6.9-inch, 120Hz Chipset Dimensity 9600 Pro (2nm) Snapdragon 8 Elite (3nm) Battery 6,510mAh 5,000mAh Charging 90W 45W Main camera 50MP 200MP Telephoto 200MP periscope 50MP periscope (5x) + 10MP (3x) Price From CNY 6,499 (China) From $1,299.99 (US), Korea price TBC On paper, vivo leads on battery, charging speed, and refresh rate across the board — 1,510mAh more capacity, nearly double the charging wattage, and a 24Hz higher refresh rate. The Galaxy S25 Ultra counters with a larger screen and a main camera rated four times higher in resolution, though that\u0026rsquo;s less a clear win than a different design choice: a 200MP main sensor is built to bin pixels together in low light to cut noise, so raw megapixel counts alone don\u0026rsquo;t settle which camera actually takes a better photo.\nTrade-offs and who each phone fits vivo X500 Pro suits buyers who prioritize battery life, fast charging, and telephoto zoom for photography. Its Log video and ACES support also make it an unusually capable pocket camera for anyone who edits footage afterward rather than shooting straight to a finished clip. It runs Funtouch OS instead of One UI, though, and has no Samsung DeX or Galaxy ecosystem ties, so the switching case is weak for anyone already invested in a Galaxy-and-Windows setup. The Galaxy S25 Ultra\u0026rsquo;s case rests on its larger screen, the S Pen, and Korea\u0026rsquo;s established retail and warranty network. It trails on battery and charging numbers, but 45W is still fast enough for most daily use in practice, and having confirmed local distribution and support is a real, practical difference vivo can\u0026rsquo;t currently match here. Buyers weighing an import purchase should also factor in the lack of official warranty coverage, carrier compatibility questions, and no guarantee of long-term software updates outside vivo\u0026rsquo;s home markets.\nAvailability in Korea vivo does not currently sell smartphones officially in South Korea, and the X500 series announcement made no mention of a Korean launch. Only the China retail date, September 29, is confirmed. Wider global availability and any Korea plans remain TBC. Anyone in Korea wanting one today would be looking at a personal import or gray-market purchase, neither of which carries the manufacturer warranty or carrier support that comes standard with a Galaxy device bought locally.\nVerdict On the spec sheet, the vivo X500 Pro beats the Galaxy S25 Ultra on battery, charging speed, and refresh rate. But the buyers who can actually feel that advantage are, for now, mostly limited to China. For anyone shopping in Korea, this isn\u0026rsquo;t really a head-to-head between two competing options — it\u0026rsquo;s a comparison between a phone you can buy and one you can only read specs about. If a 200MP main camera, a 6.9-inch screen, and the S Pen matter to you, the Galaxy S25 Ultra remains the practical pick.\n","permalink":"https://dibrain.dev/en/2026/09/22/phone-vivo-x500-pro-galaxy-s25-ultra/","summary":"vivo unveiled the X500 Pro in China on September 21. We lined its battery, charging speed and refresh rate against the Samsung Galaxy S25 Ultra, which is already on sale in Korea — vivo isn\u0026rsquo;t.","title":"vivo X500 Pro Wins on Battery, Not on Availability"},{"content":"Apple unveiled the Watch Series 12 and Ultra 4 on September 9, and both went on sale on September 18 — Korea included in the first wave, alongside a new AirPods 5 we\u0026rsquo;re not covering here. Neither watch got a redesign; the changes are concentrated in the heart-rate sensor and battery management. Since Samsung\u0026rsquo;s Galaxy Watch Ultra 2 already occupies the same \u0026ldquo;Ultra\u0026rdquo; tier, we put the numbers side by side.\nUltra 4 / Series 12 Spec Ultra 4 Series 12 Chip S11 S11 Battery (typical use) Up to 50 hours TBC Battery (Low Power Mode) Up to 84 hours - Battery (Extended Workout, continuous GPS + heart rate) 25 hours - Heart-rate sampling TBC Every 5 seconds (60x more often) Display brightness 3,000 nits TBC Size 49×44×12mm TBC Weight 63-63.1g TBC Material Titanium Aluminum / Titanium Satellite Emergency SOS None Price (US) From $799 From $399 Price (Korea) From ₩1,249,000 From ₩599,000 Ultra 4 keeps the same shell as its predecessor and puts its changes into an entirely new health-sensing system and battery management. Its price hasn\u0026rsquo;t moved either: Ultra 4 starts at $799, the same as Ultra 3, even with the overhauled sensors. The 25-hour Extended Workout figure covers continuous GPS tracking with heart-rate monitoring running the whole time — the scenario that drains a watch fastest — while the 50-hour number describes more typical mixed use, and 84 hours applies only with Low Power Mode switched on, which turns off several always-on features. Series 12\u0026rsquo;s aluminum-model front glass is 60% tougher than before, and its new heart-rate sensor samples every 5 seconds — 60 times more often than the previous generation. Both models also gain a new \u0026ldquo;readiness\u0026rdquo; score that combines activity, vital signs, and sleep data to flag whether you\u0026rsquo;re primed for a hard workout or need to recover.\nAgainst the Galaxy Watch Ultra 2 Spec Apple Watch Ultra 4 Galaxy Watch Ultra 2 Battery (typical use) Up to 50 hours Up to 60 hours Display brightness 3,000 nits 5,000 nits Processor S11 Snapdragon Wear Elite OS watchOS, iOS-only Wear OS, Android-only Size 49×44×12mm 47.4×47.1×10.7mm Weight 63-63.1g 61.5g Material Titanium Titanium Price (US) $799 $649-699 Price (Korea) ₩1,249,000 ~₩969,100 Galaxy Watch Ultra 2 beats Ultra 4 by 10 hours of battery and 2,000 nits of brightness, and it costs roughly $130-150 less in the US, about ₩280,000 less in Korea. The battery figures Apple is promoting — 50 hours, or 84 in Low Power Mode — still fall short of the competitor\u0026rsquo;s baseline number, and that\u0026rsquo;s before accounting for the fact that Apple\u0026rsquo;s 50-hour figure is itself a best-case claim for typical use rather than a guaranteed minimum.\nTrade-offs and who each watch fits Ultra 4\u0026rsquo;s case rests on satellite Emergency SOS, Apple Intelligence, and iOS integration. If you\u0026rsquo;re on an iPhone and need satellite SOS for hiking or diving where signal drops out, Ultra 4 is the pick. If battery life, outdoor visibility, and price matter more, Galaxy Watch Ultra 2 wins on all three — though it\u0026rsquo;s Android-only, so iPhone owners aren\u0026rsquo;t choosing between the two anyway. For an Android user weighing both platforms, the math is more one-sided: a longer-lasting, brighter, cheaper watch, with the only real trade-off being the loss of satellite SOS and Apple\u0026rsquo;s software ecosystem.\nSeries 12 drops the satellite feature and targets everyday wear; the 5-second heart-rate sampling is this generation\u0026rsquo;s headline change. For a daily health watch rather than an outdoor-adventure one, Series 12 makes more sense than Ultra 4. The price gap between the two Apple watches is substantial too — Ultra 4 costs $400 more than Series 12 in the US, a difference that buys the titanium case, satellite SOS, and the extended battery modes.\nAvailability in Korea Both models are already on sale in Korea as of September 18, the same day as the US launch. Pre-orders opened on September 9, right after the announcement, and Korea was included in that first wave rather than a later rollout — a pattern Apple has kept consistently for its wearables in recent years. Series 12 starts at ₩599,000, Ultra 4 at ₩1,249,000.\nVerdict This generation of Apple Watch is a health-sensor and battery-management update rather than a redesign. But the battery numbers Ultra 4 leads with don\u0026rsquo;t even match the Galaxy Watch Ultra 2\u0026rsquo;s baseline specs, and it costs more on top of that. Its 3,000-nit display also trails the competitor\u0026rsquo;s 5,000 nits, which matters most exactly when you\u0026rsquo;re outdoors checking a watch face in direct sun. Unless you need iOS lock-in or satellite SOS, this upgrade gives more reasons to stay put than to switch. And for anyone already on iPhone, the comparison is academic anyway — Galaxy Watch Ultra 2 doesn\u0026rsquo;t pair with iOS at all, so the two were never really competing for the same buyer.\n","permalink":"https://dibrain.dev/en/2026/09/21/wearable-apple-watch-ultra4-galaxy-watch-ultra2/","summary":"Apple Watch Ultra 4 and Series 12 went on sale in Korea on September 18. We lined their battery, brightness and price against the Samsung Galaxy Watch Ultra 2, already on shelves.","title":"Apple Watch Ultra 4 Costs More, Runs Out Sooner"},{"content":"Korea\u0026rsquo;s product calendar was quiet this week (Sept 11–18): we could not confirm a single new Korea-market launch or official Korea release announcement in that window, so this brief leans on global hardware and open-source software instead. Two items stood out: a screen-aware AI gadget that just opened its crowdfunding campaign at a price that will not hold, and an Alibaba-built code review tool that climbed GitHub\u0026rsquo;s trending list within days of going open source.\nKorea No confirmed Korea market launch this week. We checked manufacturer press releases and major Korean tech outlets and found nothing dated inside the Sept 11–18 window — the closest candidates (Samsung\u0026rsquo;s Galaxy S26 FE and Galaxy Book6 launches) both happened earlier in the month and were already covered in previous briefs.\nGlobal Violoop — a palm-sized AI hardware device from BVIO Technology, first shown at IFA 2026 in Berlin and now live on Kickstarter as of September 15.\nConnects to a PC or Mac via HDMI input and USB-HID: it captures the screen locally and sends keyboard/mouse commands back, effectively acting as a second operator sitting beside the machine rather than a cloud assistant reading a chat window. BVIO says the underlying model was trained on 200 apps and 10,000 real-world screens, and frames its long-term goal as \u0026ldquo;artificial intuition\u0026rdquo; — software that anticipates a task rather than waiting to be prompted for one. Early-bird price is $399, held with a refundable $10 deposit; once the campaign tier sells out, the price steps up, and retail after the campaign is $699 — a 75% jump from the opening price. Shipping is targeted for October 2026, a tight two-to-three-month window for a first hardware run. BVIO says it closed a $20 million seed round ahead of the campaign, which is unusually well-funded for a Kickstarter-stage device and suggests the crowdfunding push is more about early-adopter distribution than survival capital. Korea availability and pricing: TBC — BVIO has not announced a Korea launch or local distributor yet. Sources: PC Guide, Tom\u0026rsquo;s Guide, Violoop official site Software alibaba/open-code-review — an AI code review CLI that Alibaba says it ran internally for over two years, reviewing tens of thousands of developers\u0026rsquo; diffs, before open-sourcing it. It jumped up GitHub\u0026rsquo;s trending list this week.\nLicense: Apache-2.0 / Primary language: Go About 34.4k stars as of this writing, with roughly 8,500 of those added in the past week alone — a sign the trending-list spike, not a slow organic climb, is what\u0026rsquo;s driving the number. Latest release v1.12.5 (2026-09-17): viewer UI improvements, an optional PR-number input for CI workflows, a new Kimi Code plugin, and fixes for diff handling and token-budget enforcement. The design combines deterministic pipelines with an LLM agent: file-level static rules catch known issues first, and the agent is only invoked where judgment is needed, producing line-level comments on git diffs. Built-in rulesets cover multiple languages and common classes of bugs — null-pointer exceptions, thread-safety issues, XSS, and SQL injection. Alibaba\u0026rsquo;s own benchmark claims, per the project README, are that Open Code Review reaches higher precision and F1 than general-purpose coding agents (it names Claude Code specifically) on the same underlying model, while using roughly one-ninth of the tokens — a trade-off the maintainers describe as favoring precision over recall, meaning it flags fewer issues overall but is more often right about the ones it does flag. Treat this as the vendor\u0026rsquo;s own claim rather than an independent benchmark. Install and run: npm install -g @alibaba-group/open-code-review ocr config provider ocr config model Requires Git 2.41 or later. The ocr command becomes available globally after install, and it integrates with Claude Code, Codex, Cursor, and Kimi Code as plugins alongside standalone CLI and CI use. Source: github.com/alibaba/open-code-review This Week at a Glance Section Item Key number Korea — No confirmed launch Global Violoop $399 early bird / $699 retail Software alibaba/open-code-review 34.4k stars, +8.5k this week A quiet Korea week doesn\u0026rsquo;t mean a quiet week overall — it just means the interesting movement happened in crowdfunding pricing and open-source adoption curves instead of retail shelves. Both Violoop and Open Code Review are worth a second look once they have a track record: Violoop once early units ship and reviewers can test the HDMI-capture latency claim in practice, and Open Code Review once its precision/recall trade-off has been tested outside Alibaba\u0026rsquo;s own benchmark.\nThis Week on DigitalBrain iPhone Duo Folds Late, Costs More Nurovi\u0026rsquo;s UV Light Meets Aqua20\u0026rsquo;s 180°C Steam UR Gen 7: Quicker Arm, Smaller Brain MiniCPM5 Wins Big, Flash-Next Spends Small ","permalink":"https://dibrain.dev/en/2026/09/18/brief-violoop-open-code-review/","summary":"No confirmed Korea product launches this week. Globally, screen-watching AI hardware Violoop opened its Kickstarter at $399, and on the software side Alibaba\u0026rsquo;s open-sourced code review tool gained about 8,500 GitHub stars in a single week.","title":"Violoop's $399 Won't Last the Campaign"},{"content":"Two picks from this week\u0026rsquo;s Hugging Face trending list take opposite routes to the same goal. One shrinks the whole model to win. The other keeps the total size but only wakes up a fraction of it per token. Both target the same practical problem — running a capable model without sending every request to someone else\u0026rsquo;s cloud — but they solve it at completely different scales. If you want something to run on your own hardware, both are worth a look.\nMiniCPM5-2B Model card: openbmb/MiniCPM5-2B Source: GitHub - OpenBMB/MiniCPM Publisher: OpenBMB License: Apache-2.0 Parameters: 2.52B dense, 42 layers, GQA (16 query heads / 2 KV heads) Context: 131,072 tokens Released: 2026-09-07 The second checkpoint in the MiniCPM5 line, built for local and on-device use rather than cloud serving. Its grouped-query attention setup — 16 query heads sharing just 2 key/value heads — is one of the reasons it stays light enough for phone-class memory budgets while still handling a 131K-token context. OpenBMB\u0026rsquo;s own numbers put its average across 34 benchmarks at 53.9, ahead of the larger Qwen3.5-4B at 51.1, despite having roughly half the parameters. Standouts include MATH-500 at 94.6, AIME 2026 at 86.5, LiveCodeBench v6 at 69.1, SWE-bench Verified at 46.4, τ²-Bench Telecom at 97.1, and GAIA Text-103 at 88.7 (all OpenBMB\u0026rsquo;s own figures). It shipped same-day in GGUF, MLX, GPTQ, and mobile-oriented LiteRT builds, so it\u0026rsquo;s ready to try on a laptop or a phone-class device right away, without waiting on a community quantization pass.\nRun example (transformers):\nfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = \u0026#34;openbmb/MiniCPM5-2B\u0026#34; tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=\u0026#34;auto\u0026#34;, device_map=\u0026#34;auto\u0026#34; ) messages = [{\u0026#34;role\u0026#34;: \u0026#34;user\u0026#34;, \u0026#34;content\u0026#34;: \u0026#34;Who are you?\u0026#34;}] inputs = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, enable_thinking=True, return_dict=True, return_tensors=\u0026#34;pt\u0026#34; ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=128) print(tokenizer.decode(outputs[0][inputs[\u0026#34;input_ids\u0026#34;].shape[-1]:])) It also runs out of the box on Ollama, llama.cpp, LM Studio, and vLLM. Good for: developers who want to run coding, math, or agentic workloads locally without server costs, or teams building an on-device chatbot.\nQwen3.8-Flash-Next Model card: Qwen/Qwen3.8-Flash-Next GGUF: unsloth/Qwen3.8-Flash-Next-GGUF Publisher: Qwen Team, Alibaba Group License: Qwen Community License 1.0 (services over 100M monthly active users or $20M monthly revenue, and Model-as-a-Service or AI work-assistant businesses, need a separate license — not Apache-2.0) Parameters: sources disagree here — Qwen\u0026rsquo;s own material describes 125B total with 6B active per token (512 experts, 10 routed + 1 shared), plus a separate 51B n-gram embedding and a 4B MTP layer; some third-party write-ups add these up to a 177B total. TBC. Context: 262,144 tokens Task: multimodal (text + vision) MoE Released: 2026-08-26 Released as the first open-weight preview of the architecture behind Qwen4. The hybrid design repeats three Gated DeltaNet layers plus one Qwen Sparse Attention (QSA) block across the network; QSA operates at the micro-block level rather than per token, which is what lets the 262K-token context stay usable for agentic workloads instead of falling off a latency cliff. The sparse MoE routing — 512 experts with 10 routed plus 1 shared expert active at a time — is the whole point: only 6B of the total footprint actually does compute for any given token, even though the full checkpoint has to sit in memory. Community GGUF quantizations run in roughly 82GB at 4-bit (58GB for the main model plus 24GB for the n-gram component), so 75GB or more of unified memory is enough with no discrete GPU required. One community report has an 85GB quant running on an M5 Max Mac with 64GB of memory at 36 tokens/sec with vision input (third-party benchmark, not an official figure). Treat this as workstation-class on-device, not phone-class — it needs serious unified memory or a multi-GPU box, just not a datacenter.\nRun example (vLLM):\nvllm serve Qwen/Qwen3.8-Flash-Next --port 8000 --tensor-parallel-size 4 GGUF builds load directly in llama.cpp. Good for: developers with a high-memory unified Mac or a multi-GPU box who want an early look at the Qwen4-generation architecture.\nWrap-up Both models chase \u0026ldquo;on-device\u0026rdquo; from opposite directions. MiniCPM5-2B shrinks the parameter count itself down to 2.52B, low enough to reach phones. Qwen3.8-Flash-Next keeps a 125B total but caps active parameters per token at 6B, aiming at high-memory Macs and workstations instead. Need something light to run locally today? Go with MiniCPM5-2B. Want a preview of where the next architecture generation is headed? Qwen3.8-Flash-Next is the one to try.\n","permalink":"https://dibrain.dev/en/2026/09/17/dev-minicpm5-qwen-flash-next/","summary":"This week\u0026rsquo;s Hugging Face trending picks: MiniCPM5-2B, a 2.52B on-device model that beats a 4B rival, and Qwen3.8-Flash-Next, a 125B MoE that only wakes up 6B parameters per token.","title":"MiniCPM5 Wins Big, Flash-Next Spends Small"},{"content":"Universal Robots unveiled its seventh-generation collaborative robot platform, \u0026ldquo;UR Gen 7,\u0026rdquo; on September 14 at IMTS 2026 in Chicago, one of North America\u0026rsquo;s largest manufacturing trade shows. Three new arms arrive alongside a new controller and a redesigned tool flange, and the headline change is that cameras and force sensors can now plug straight into the wrist of the arm. Universal Robots describes it as an end-to-end redesign built on 20 years of cobot work and more than 100,000 industrial deployments, and frames the launch around \u0026ldquo;physical AI\u0026rdquo; — robots that pair vision and touch sensing with real-time control rather than just repeating a fixed motion path.\nThat framing matters for how to read this generation. Instead of one flagship arm, Universal Robots shipped three variants that trade payload for reach, all sharing the same controller and flange architecture. That lets an integrator pick the geometry a task needs — longer reach for a wide work cell, or more payload for a heavier tool — without switching software or rewiring the sensor stack.\nUR Gen 7 The three new g-Series arms carry their payload (kg) and reach (mm) right in the model name.\nModel Payload Reach Weight Repeatability Max TCP speed UR10g-1750 8 kg (10 kg extended) 1750 mm 44.7 kg ±0.08 mm 5 m/s UR17g-1300 15 kg (17.5 kg extended) 1300 mm 40.7 kg ±0.05 mm 5 m/s UR18g-950 18 kg 950 mm 39.2 kg ±0.05 mm 4 m/s The controller has been replaced by the \u0026ldquo;CB7 Core,\u0026rdquo; which Universal Robots says delivers 40% more compute in a 30% smaller footprint than the previous generation. That extra compute has to go somewhere: the company positions it as headroom for running vision and force-sensing models locally, close to the arm, rather than shipping raw sensor data to an external PC.\nFrom an embedded-systems angle, the more interesting change is the new g-Series tool flange. It routes data, power, and safety signals to the end effector through a single connector at the wrist, so cameras and force sensors no longer need cabling run externally along the arm — a detail that matters a lot once you\u0026rsquo;ve had to route a MIPI or USB3 camera cable through a moving joint without it fraying. The arms themselves now ship with built-in force-torque sensing, impedance control, and Real-Time Data Exchange (RTDE), aimed at contact-sensitive tasks like precision assembly, where the arm needs to feel resistance and back off rather than push through it. The launch materials don\u0026rsquo;t specify the exact voltage/current range the tool flange supplies or each model\u0026rsquo;s peak power draw — those remain TBC, and worth confirming before designing a cabinet or power budget around one.\nUR17g-1300 vs. UR10e — same 1300 mm reach, different generation We lined up the UR17g-1300 against UR10e, the previous generation\u0026rsquo;s flagship at the same 1300 mm reach.\nSpec UR10e (previous gen) UR17g-1300 (Gen 7) Payload 12.5 kg 15 kg (17.5 kg extended) Reach 1300 mm 1300 mm Weight 33.5 kg 40.7 kg Repeatability ±0.05 mm ±0.05 mm Max TCP speed 4 m/s 5 m/s Max power draw 615 W TBC At the same reach, payload rises 20% and top speed rises 25%, but the arm itself gets more than 7 kg heavier. That extra mass looks like the price of the new tool flange and built-in force-torque sensing; repeatability holds steady at ±0.05 mm.\nWho should care Lines that bolt multiple cameras or force sensors onto the wrist — precision assembly, inspection, bin-picking with vision feedback — stand to gain the most from the simpler cabling the tool flange enables, and from having force-torque sensing built in rather than bolted on as a separate module. For a plain pick-and-place line with little sensing, the existing e-Series arms already do the job, and there\u0026rsquo;s little reason to move to a heavier, likely pricier Gen 7 arm just for the extra controller headroom. Integrators already running RTDE-based force control on e-Series arms should find the migration path familiar, since the underlying control concepts carry over — it\u0026rsquo;s the flange and controller hardware that changed, not the programming model. Pricing wasn\u0026rsquo;t disclosed at launch.\nAvailability in Korea Universal Robots runs an official distributor network in Korea, but neither the launch materials nor any local distributor notice mentions a Korea launch date or price for UR Gen 7 yet — TBC.\nVerdict UR Gen 7 changes the tool flange and controller more than it changes the arm itself. The emphasis on simpler wiring and easier sensor mounting points toward \u0026ldquo;physical AI\u0026rdquo; workloads that lean heavily on cameras and force sensing. The numbers that matter most for actually specifying a line — power draw and Korea pricing — are still missing, so anyone evaluating a purchase should wait for the full spec sheet.\n","permalink":"https://dibrain.dev/en/2026/09/16/embedded-universal-robots-gen7/","summary":"Universal Robots unveiled its seventh-generation cobot platform, UR Gen 7, at IMTS 2026. We compare the UR17g-1300 to the previous-generation UR10e, which shares the same 1300 mm reach.","title":"UR Gen 7: Quicker Arm, Smaller Brain"},{"content":"Dyson and Dreame both used IFA 2026 (September 4–8, Berlin) to reveal new premium robot vacuums. Dyson brought \u0026ldquo;Nurovi,\u0026rdquo; a three-model lineup that uses UV and green-light sensors to spot stains invisible to the naked eye. Dreame countered with the Aqua20 Pro Ultra Roller X Complete, which pairs 180°C steam with an extending roller mop. Both are all-in-one suction-plus-mop machines, so they land in direct competition, and both are shipping within days of each other in September 2026 rather than sitting on a roadmap.\nIt\u0026rsquo;s a useful pairing because the two companies are chasing the same problem — dried-on stains that a plain mop pass leaves behind — from opposite directions. Dyson is spending its engineering budget on finding the stain; Dreame is spending it on destroying the stain once found. That split shows up clearly once you line up the actual hardware.\nDyson Nurovi Dyson says the Nurovi name blends its new Neural Intelligence tech with its latest vision systems. The lineup ships in three tiers.\nModel Suction Key feature Price Availability R3 Nurovi Spot+Scrub UV (flagship) 30,000Pa UV and green-light sensors detect 200+ stain types; a 12-point moistening system uses water heated to 158°F (~70°C) $1,199 US, September 15 R2 Nurovi Wash+Dry TBC Hybrid vacuum-and-mop, Costco exclusive TBC September 3 (Costco) R1 Nurovi Dry 21,000Pa Vacuum-only, Twin-capture debris system TBC Later in September Dyson claims the R3\u0026rsquo;s suction is 65% higher than its existing Spot+Scrub Ai model.\nDreame Aqua20 Pro Ultra Roller X Complete Suction: 40,000Pa, paired with the \u0026ldquo;HyperStream Detangling Floor Brush 2.0\u0026rdquo; Steam and mop: Four outlets on the body release steam up to 180°C directly onto the floor to loosen stains, which a heated-water roller mop then wipes away Roller extension: The \u0026ldquo;MopExtend\u0026rdquo; roller can push out up to 8cm to reach edges and under furniture, spinning at 300rpm with 18N of downward pressure Navigation: \u0026ldquo;OmniSense 4.0\u0026rdquo; (dual X-Vision AI cameras plus VersaLink navigation) recognizes over 320 object types as small as 1cm within 0.1 seconds, including pet waste avoidance Obstacle clearance: \u0026ldquo;ProLeap\u0026rdquo; tech lets it climb thresholds up to 10cm Smart home: Matter-certified Price and launch: €1,499 in Germany, Austria, Spain, and Portugal; £1,299 in the UK; launching September 21. US pricing is TBC Head to head Spec Dyson R3 Nurovi Dreame Aqua20 Pro Ultra Roller X Suction 30,000Pa 40,000Pa Stain removal UV/green-light detection + 158°F heated-water mop 180°C direct steam + heated-water roller mop Threshold climbing TBC Up to 10cm Smart home TBC Matter-certified Price $1,199 (US) €1,499 / £1,299 Launch September 15 (US) September 21 (Europe) Suction is the one metric with matching units and a clear \u0026ldquo;higher is better\u0026rdquo; reading, so it\u0026rsquo;s directly comparable. The rest of the specs aren\u0026rsquo;t confirmed on both sides yet, so we kept this as a table rather than a bar chart.\nTrade-offs Dyson Nurovi\u0026rsquo;s strength is detection precision — its UV and green-light sensors can flag stains people would never spot on their own — and at $1,199 for the flagship, the entry price is relatively low for the category. Its weakness: 30,000Pa trails the competition, and a heated-water mop (rather than steam) may struggle with baked-on grease.\nDreame\u0026rsquo;s Aqua20 wins on raw cleaning power, with 40,000Pa suction and 180°C steam that should cut through stubborn stains, plus an 8cm-extending roller for better edge coverage. The trade-offs are a higher price (€1,499) and no confirmed US or Korea pricing or launch date yet.\nIf you want a robot that finds pet hair and faint stains with precision, Nurovi is the pick. If you want raw steam-powered cleaning force for baked-on grime, Aqua20 wins.\nNeither company has published independent lab results yet, so these are still manufacturer-stated numbers. Suction figures in particular tend to be measured under best-case lab conditions rather than real carpet or hard-floor resistance, so treat the 30,000Pa-vs-40,000Pa gap as a starting point for comparison, not a guarantee of how much dirt actually comes up on your floor.\nMaintenance is worth weighing too, even though neither company has published full cost breakdowns yet. Steam systems like Aqua20\u0026rsquo;s need clean water and periodic descaling to keep the internal heater working, and a 180°C steam line is one more component that can fail outside warranty. Dyson\u0026rsquo;s heated-water approach is mechanically simpler, which may mean fewer parts to maintain, but it also means less thermal energy hitting the stain in the first place. Both are trade-offs worth confirming against real owner reports once these units have been in homes for a few months, not just spec sheets from launch day.\nKorea availability As of this writing, neither product has a confirmed Korean launch date (TBC). Nurovi isn\u0026rsquo;t yet listed on Dyson Korea\u0026rsquo;s robot vacuum page. Dreame\u0026rsquo;s previous-generation Aqua10 Ultra Roller sells in Korea for roughly ₩890,000–940,000, but pricing and availability for the new Aqua20 Pro Ultra Roller X Complete remain unconfirmed.\nVerdict Both companies are making real performance claims, but they\u0026rsquo;re betting on different strengths: Dyson is betting on better stain detection, Dreame on stronger stain removal. On paper, Aqua20\u0026rsquo;s 40,000Pa suction edges out Nurovi\u0026rsquo;s 30,000Pa, but real-world differences may show up more in detection accuracy and maintenance costs (consumables, water tank upkeep) than in raw suction numbers. With Korean pricing and launch dates still unconfirmed for both, it\u0026rsquo;s more sensible to wait for official announcements than to weigh a purchase right now.\n","permalink":"https://dibrain.dev/en/2026/09/15/appliance-dyson-nurovi-dreame-aqua20/","summary":"Dyson unveiled its \u0026lsquo;Nurovi\u0026rsquo; robot vacuum lineup at IFA 2026, while Dreame showed off the Aqua20 Pro Ultra Roller X Complete at the same show. We compare suction power, stain removal, and navigation.","title":"Nurovi's UV Light Meets Aqua20's 180°C Steam"},{"content":"Apple unveiled its first foldable iPhone, the iPhone Duo, at its \u0026ldquo;Surprise and Shine\u0026rdquo; event on September 9. Announced alongside the iPhone 18 Pro lineup, it\u0026rsquo;s Apple\u0026rsquo;s long-awaited answer to a foldable category that Samsung has already iterated through eight generations of its Galaxy Z Fold line, putting the two devices under immediate comparison. Samsung shipped the original Galaxy Fold back in 2019, so this is Apple\u0026rsquo;s response seven years into the category — which makes the \u0026ldquo;why now, why this approach\u0026rdquo; question as interesting as the spec sheet itself.\niPhone Duo Spec Detail Inner display 7.6-inch OLED, ProMotion, always-on Outer display 5.4-inch OLED, ProMotion, always-on Peak outdoor brightness 3,000 nits Frame Titanium Water/dust resistance IP68 Chip TBC US price (256GB) $1,999 Korea price (256GB) ₩3.29 million The chip is listed as TBC (to be confirmed) because, while the iPhone 18 Pro announced at the same event reportedly ships with the 2nm A20 Pro, Apple has not published an official spec sheet confirming the same chip inside the iPhone Duo. (Source: MacRumors)\nBoth the inner and outer panels support ProMotion\u0026rsquo;s variable refresh rate and always-on display, which reads as Apple trying to preserve the standard iPhone screen experience even on a folding form factor. What wasn\u0026rsquo;t detailed at the event is whether the inner display shows a visible crease, or how many open-close cycles the hinge is rated to survive — durability questions that typically only get answered once review units ship.\niPhone Duo vs. Galaxy Z Fold8 Spec iPhone Duo Galaxy Z Fold8 Inner display 7.6-inch 7.6-inch Outer display 5.4-inch 5.5-inch Frame Titanium Aluminum Water/dust resistance IP68 IP48 Weight TBC 53g lighter than iPhone Duo Thickness (folded/unfolded) TBC Thinner than iPhone Duo in both states US price (256GB) $1,999 $1,899 On paper, the two foldables sit in the same weight class for screen size: both have 7.6-inch inner displays, and the outer screens are nearly identical at 5.4 and 5.5 inches. The real split is in materials and durability. Apple went with a titanium frame and an IP68 rating, while Samsung kept its aluminum frame and IP48 rating in exchange for a lighter, thinner device. The iPhone Duo also costs $100 more at launch.\nBuyers who prioritize water resistance and build durability have a clear case for the iPhone Duo. Those who care more about portability and weight are better served by the Galaxy Z Fold8. That said, a number of the iPhone Duo\u0026rsquo;s specs — including the chip — remain unconfirmed on an official spec sheet, so a real performance comparison will have to wait until after launch. (Source: TechRadar, GSMArena)\nZoom out to the category itself, and Samsung\u0026rsquo;s eighth generation brings years of accumulated hinge-engineering and software-optimization experience that a first-generation product can\u0026rsquo;t match overnight. Apple, on the other hand, is carrying over split-screen and multitasking features built for iPadOS onto a folding phone screen for the first time — how naturally those translate in daily use is the kind of thing that\u0026rsquo;s hard to judge from a spec sheet alone, and will really only be settled once review units are in hand.\nKorea Availability South Korea is among the iPhone Duo\u0026rsquo;s first-wave launch markets. Pre-orders open at 9pm KST on October 16, with the official on-sale date set for October 23. Korean pricing runs ₩3.29 million for 256GB, ₩3.59 million for 512GB, ₩4.19 million for 1TB, and ₩5.09 million for 2TB. A straight currency conversion of the $1,999 US price for the 256GB model comes out to roughly ₩2.67 million, meaning the Korean price is priced nearly ₩600,000 higher. While the standard iPhone 18 Pro lineup ships in 65 countries starting September 18, the foldable Duo follows on a separate schedule more than a month later. (Source: ZDNet Korea, Asia Economy)\nThat month-long gap is probably not incidental. Unlike the standard iPhone, which Apple has mass-produced annually for over a decade, the foldable is a brand-new form factor to manufacture at scale, and hinge-assembly yield issues would plausibly need extra time to iron out before a wider international rollout. Carrier pre-order promotions and subsidy amounts in Korea haven\u0026rsquo;t been announced yet, so the actual out-of-pocket price will need to be confirmed once carriers publish their offers.\nVerdict Apple\u0026rsquo;s first foldable has caught up to the Galaxy Z Fold8 in screen size, but that parity comes at a higher price than Samsung\u0026rsquo;s. It leads on water resistance and materials, but trails on weight and thickness. Coming late to the foldable market, Apple appears to have chosen a \u0026ldquo;more expensive, more durable\u0026rdquo; position — whether that positioning works will become clearer once Korean sales figures come in after the October 23 launch.\nThe bigger question is whether Apple\u0026rsquo;s brand alone is enough to expand a foldable category that, despite eight generations from Samsung, still hasn\u0026rsquo;t broken into the mainstream the way standard smartphones have. If the iPhone Duo pulls first-time foldable buyers away from flat iPhones rather than converting existing Android foldable owners, that would mark a bigger shift in the market than anything in the spec comparison above — and it\u0026rsquo;s a question that only sales data, not a launch keynote, can answer.\n","permalink":"https://dibrain.dev/en/2026/09/14/phone-iphone-duo-galaxy-fold8/","summary":"Apple unveiled its first foldable phone, the iPhone Duo, at its September 9 event alongside the iPhone 18 Pro lineup. Here\u0026rsquo;s how it compares with the Galaxy Z Fold8 on specs and price, plus Korea\u0026rsquo;s launch timeline.","title":"iPhone Duo Folds Late, Costs More"},{"content":"At IFA 2026, Acer unveiled the Swift Blade 14, pitched as the world\u0026rsquo;s lightest 14-inch laptop at 799g — a number aimed squarely at LG\u0026rsquo;s \u0026ldquo;gram\u0026rdquo; line, the brand that has owned the ultralight-laptop label for years. The timing is pointed: LG just launched its own gram AI 2026 14-inch in Korea earlier this month, so the two are worth putting side by side.\nOn-device AI is this year\u0026rsquo;s marketing headline for laptops, but what actually decides a purchase is still weight and battery — the two specs buyers feel every single day, long after an NPU benchmark number has stopped mattering. The Swift Blade 14 reignites that more basic fight. Acer also announced a 16-inch sibling at the same event, the Swift Air 16 (13.3mm thick, 1.5kg, 3K OLED, 70Wh battery, and 100W fast charging that hits 50% in 30 minutes), but the number that matters here is the one the 14-inch model set: under 800 grams.\nAcer Swift Blade 14 Spec Value Weight 799g Thickness 12.9mm Display 14\u0026quot; WQXGA+ OLED (2,880×1,800), 90Hz Processor Up to Intel Core 7 350 (integrated graphics) Memory / storage 12GB LPDDR5X, up to 1TB SSD Battery 50Wh Chassis Magnesium-aluminum alloy + carbon fiber Ports 3× USB 3.2 Gen 1 Type-C, FHD webcam with physical shutter Colors Marshmallow Blue / Marshmallow White Price Under €2,000 in EMEA (exact pricing TBC) Availability EMEA December 2026, North America Q1 2027 (Source: Acer press release, fonearena)\nThe chassis is magnesium-aluminum alloy combined with carbon fiber, a materials choice ultralight laptops typically lean on to hit sub-kilogram weights while keeping an OLED panel. Acer hasn\u0026rsquo;t published NPU or on-device-AI performance figures for the Core 7 350, so a direct AI-performance comparison with the gram\u0026rsquo;s Core Series 3 chip stays unconfirmed for now (TBC).\nCompetitor comparison — LG gram AI 2026 14\u0026quot; Spec Acer Swift Blade 14 LG gram AI 2026 14\u0026quot; Weight 799g 1.29kg Thickness 12.9mm 14.9mm Battery 50Wh 61.2Wh (LG claims up to 30.5 hours) Processor Intel Core 7 350 2026 Intel Core Series 3 Price Under €2,000 (EMEA) ₩1.45M (Korea list price) Korea launch TBC Already launched (Sep 7, 2026) (Source: betanews on the LG gram AI 2026, NGO News Korea)\nThe Swift Blade 14 undercuts the gram by roughly 500g and is 2mm thinner, but it also carries a smaller battery. Acer hasn\u0026rsquo;t published official battery-life figures, so a direct runtime comparison isn\u0026rsquo;t possible yet (TBC). On the display side the Swift Blade 14\u0026rsquo;s 90Hz OLED looks like the stronger spec, but LG hasn\u0026rsquo;t detailed the gram AI 2026\u0026rsquo;s panel, so that comparison is incomplete too.\nTrade-offs and who it\u0026rsquo;s for The appeal is straightforward: dropping under 800g means you can throw in a power bank or a portable monitor without the bag suddenly feeling heavy, and 90Hz OLED is rare in this weight class. The webcam\u0026rsquo;s physical privacy shutter is also a genuinely useful touch for anyone doing frequent video calls in hybrid work. The downside is the 50Wh battery and just three USB-C ports — road warriors who need all-day runtime or plug in a lot of peripherals may find the gram\u0026rsquo;s 61.2Wh pack more forgiving, and an all-USB-C port layout means a dongle or hub is mandatory if you still carry legacy peripherals. If weight is the single deciding factor and you travel constantly, the Swift Blade 14 wins; if you need to buy today with local support, the gram AI 2026 is the practical pick.\nKorea availability Acer Korea has not announced a launch date or local pricing for the Swift Blade 14. TBC. There is a precedent worth noting, though: back in April, Acer brought its 0.99kg ultralight Swift Edge 14 AI to Korea through a Coupang pre-order campaign priced around ₩1.39M. A Korea launch following the EMEA and North America rollout is plausible, but both timing and pricing remain unconfirmed until Acer Korea makes an official announcement.\nVerdict On paper, the Swift Blade 14\u0026rsquo;s weight undercuts the very name \u0026ldquo;gram,\u0026rdquo; but without confirmed battery life or a Korea launch date, this is still half an announcement. The 12GB of RAM also stands out as modest next to the gram\u0026rsquo;s platform, at least based on what Acer has disclosed so far — whether that figure holds through to the shipping configuration is something only the December launch will confirm.\nFor now, the Swift Blade 14 is best read as a statement of intent: Acer wants the lightest-14-inch-laptop claim, and 799g is a real number backed by an official spec sheet, not a rumor. Whether it holds up as a buying decision depends on three things Acer hasn\u0026rsquo;t confirmed yet — actual battery runtime under normal use, final EMEA and North America pricing, and whether Korea gets it at all. Once EMEA sales begin in December and real reviews confirm battery life and final pricing, it\u0026rsquo;s worth a second look. Until then, buyers who need an ultralight 14-inch machine today have a known quantity in the gram AI 2026, already on shelves in Korea at a confirmed ₩1.45M.\n","permalink":"https://dibrain.dev/en/2026/09/11/laptop-acer-swift-blade-14/","summary":"Acer unveiled the 799g Swift Blade 14 at IFA 2026. We weigh it against LG\u0026rsquo;s 1.29kg gram AI 2026 14-inch, which just launched in Korea this month.","title":"Lighter Than a Gram: Acer's Swift Blade 14"},{"content":"This week\u0026rsquo;s GitHub trending page is really about what you install for an agent, not as an agent. One repo is pure instructions — no application code — and still crossed a quarter-million stars. The other is a browser-automation MCP server that just shipped a plugin layer. Both install in one npx line, and both lean on the same idea that\u0026rsquo;s been spreading across the agent ecosystem this year: package an agent\u0026rsquo;s behavior the way you\u0026rsquo;d package a library, then let a one-line installer pull it into whatever coding assistant you happen to be running.\nmattpocock/skills Matt Pocock, known for his TypeScript teaching, publishes this repo of agent skills — folders of instructions and scripts that AI coding agents read and follow, rather than application code. At the time of writing it sits at 257,729 stars, having gained 13,419 stars this week alone. It\u0026rsquo;s MIT-licensed; the primary distribution mechanism is a shell-based installer. No release tag was visible on the repo page, so that detail is marked TBC.\nAccording to the README, the skills are built on real engineering experience: small, composable tools that get developers and AI agents speaking the same vocabulary, tighten code quality through feedback loops, and keep architecture consistent. They\u0026rsquo;re pitched as model-agnostic — usable with any AI model, not tied to one vendor.\nInstallation is the one-liner the README leads with:\n# As a Claude Code plugin claude plugins install mattpocock-skills # or, inside a session /plugin install mattpocock-skills # For other agents/tools npx skills@latest add mattpocock/skills # After installing /setup-matt-pocock-skills Who it\u0026rsquo;s for: senior developers who use Claude Code or Codex daily and are tired of re-explaining code-review conventions, naming, and architecture rules in every prompt.\nWhat makes the repo worth watching is less any single skill and more the packaging format itself. A skill is just a folder — a markdown file of instructions plus, optionally, scripts and reference material — that an agent can discover and load on demand instead of having the whole thing stuffed into a system prompt. That\u0026rsquo;s a small technical idea, but distributing it through a package-manager- style installer, with claude plugins install on one side and a generic npx skills@latest add on the other, is what let one person\u0026rsquo;s personal conventions grow into a quarter-million-star repository overall, with this week\u0026rsquo;s trending data alone showing another 13,419-star jump on top of that.\nSource: github.com/mattpocock/skills ChromeDevTools/chrome-devtools-mcp Built by the Chrome DevTools team, this is an MCP (Model Context Protocol) server that lets AI coding agents open a real Chrome browser to get performance traces, network and console debugging, and Puppeteer-based automation. It\u0026rsquo;s written in TypeScript under an Apache-2.0 license, with 51,455 stars at the time of writing (+1,014 this week).\nThe latest release, v1.9.0, shipped on September 8, 2026 — the day before this was checked. The release notes list an \u0026ldquo;Agent Plugins 1.0 package\u0026rdquo; and a new option to disable JavaScript-execution tools. Where Matt Pocock\u0026rsquo;s repo installs instructions, this release opens an extension slot on the tooling side.\nAdd it to your MCP client config:\n{ \u0026#34;mcpServers\u0026#34;: { \u0026#34;chrome-devtools\u0026#34;: { \u0026#34;command\u0026#34;: \u0026#34;npx\u0026#34;, \u0026#34;args\u0026#34;: [\u0026#34;-y\u0026#34;, \u0026#34;chrome-devtools-mcp@latest\u0026#34;] } } } A lighter, headless mode is also available:\n{ \u0026#34;mcpServers\u0026#34;: { \u0026#34;chrome-devtools\u0026#34;: { \u0026#34;command\u0026#34;: \u0026#34;npx\u0026#34;, \u0026#34;args\u0026#34;: [\u0026#34;-y\u0026#34;, \u0026#34;chrome-devtools-mcp@latest\u0026#34;, \u0026#34;--slim\u0026#34;, \u0026#34;--headless\u0026#34;] } } } Who it\u0026rsquo;s for: frontend and QA engineers who want an agent to open a real web app and debug performance or network issues on its own, instead of pasting console logs back and forth by hand.\nThe two config blocks above are the whole setup — no build step, no separate binary to manage, because npx -y chrome-devtools-mcp@latest always pulls the current published version. The --slim --headless flags matter for CI or for running the server on a machine without a display: slim mode trims the tool surface down to the essentials so the agent isn\u0026rsquo;t offered debugging commands it can\u0026rsquo;t usefully call in a headless context.\nSource: github.com/ChromeDevTools/chrome-devtools-mcp, releases page Wrap-up Both READMEs lead with a single install command — one for instructions, one for tooling — and both landed on trending the same week. It\u0026rsquo;s worth noting this isn\u0026rsquo;t happening in just one corner of the ecosystem: OpenAI runs its own openai/skills catalog for Codex (MIT-style per-skill licensing, an skill-installer CLI, though that particular repo now points users to a newer OpenAI Plugins repository instead). Anthropic, OpenAI, and now a browser-tooling project maintained by the Chrome DevTools team are all converging on the same shape: package agent behavior — instructions or tools — as an installable unit, and let a one-line command wire it into whichever assistant a developer happens to run that day.\nRepo Stars (checked) Gained this week License mattpocock/skills 257,729 +13,419 MIT ChromeDevTools/chrome-devtools-mcp 51,455 +1,014 Apache-2.0 ","permalink":"https://dibrain.dev/en/2026/09/10/dev-mattpocock-skills-chrome-devtools/","summary":"This week\u0026rsquo;s GitHub trending list pairs Matt Pocock\u0026rsquo;s agent-skills repo, which crossed a quarter-million stars, with Chrome DevTools MCP v1.9.0, which just added a plugin system for agents.","title":"Matt Pocock's Skills: One Line, No Code"},{"content":"CNX Software covered YuzukiNeko on September 8, and it\u0026rsquo;s an unusual board: same 2x20-pin form factor as a Raspberry Pi Pico, but instead of a microcontroller it carries an application processor that runs Linux. Pico-sized boards have so far been strictly microcontroller territory. This one breaks that pattern and adds display and video capability on top.\nYuzukiNeko Spec Detail SoC Allwinner F101S3, single XuanTie C907 RISC-V core Clock 1.008GHz Memory 16MiB PSRAM Storage 16MiB external NOR flash (XIP-capable) + microSD slot Connectors USB-C, FEL button, two 2x20-pin Pico-style headers Display RGB888, LVDS, 4-lane MIPI DSI Video JPEG/PNG decode, JPEG/MJPEG encode Audio DAC, I2S/PCM, OWA output OS Linux, Zephyr RTOS License Hardware design released under CC0 1.0 (public domain) Price / availability TBC — not yet on sale (TBC) The F101S3 SoC itself was designed as a \u0026ldquo;smart control and display processor\u0026rdquo; aimed at display and audio workloads. Pairing it with PSRAM and NOR flash, then booting Linux on it, is what turns a Pico-shaped board into a real application processor rather than a microcontroller.\nThe CPU core is worth a second look, too. XuanTie C907 comes from T-Head, the chip design arm behind several of the RISC-V cores already shipping in Allwinner\u0026rsquo;s lineup. Putting one into a board this small, at this price class, is part of a broader pattern: RISC-V application cores are trickling down from mid-range SoCs into hobbyist form factors that used to be exclusively Arm or microcontroller-only. YuzukiNeko is a small but concrete data point for that trend, not a one-off curiosity.\nThe embedded angle Memory ceiling: 16MiB of PSRAM is tight for Linux. This is Buildroot- or OpenWRT-class minimal-userland territory, not a desktop distro. Power: USB-C is confirmed as the power connector, but the exact input voltage and current range aren\u0026rsquo;t published (TBC). No thermal figures are available yet either. Connectors: two 2x20-pin headers raise the possibility of drop-in compatibility with existing Pico carrier boards, but whether the pinout actually matches RP2040/RP2350 is not officially confirmed. On a robot or piece of equipment: MIPI DSI and video decode make it plausible for a small status display or a simple camera preview, but for deterministic real-time control, a microcontroller still beats a single-core Linux board. YuzukiNeko vs Raspberry Pi Pico 2 Same footprint, very different character once you put the two side by side.\nSpec YuzukiNeko Raspberry Pi Pico 2 Form factor Pico-style 2x20-pin Pico-style 2x20-pin CPU 1x XuanTie C907 RISC-V, 1.008GHz 2x Cortex-M33 + 2x Hazard3 RISC-V, 150MHz Memory 16MiB PSRAM 520KiB on-chip SRAM Onboard storage 16MiB external NOR flash 4MB onboard flash Display output RGB888 / LVDS / MIPI DSI None Runtime Linux, Zephyr RTOS Bare metal, MicroPython, CircuitPython Price TBC $5 (confirmed, always in stock) By raw capacity, YuzukiNeko dwarfs the Pico 2, but the memory types aren\u0026rsquo;t the same (PSRAM vs. on-chip SRAM). And the Pico 2\u0026rsquo;s 150MHz dual-core setup is a microcontroller built for deterministic real-time control, while YuzukiNeko\u0026rsquo;s 1.008GHz single core exists to run Linux — clock speed alone doesn\u0026rsquo;t make one \u0026ldquo;faster\u0026rdquo; than the other. These are different classes of board doing different jobs.\nPros, cons, and who it\u0026rsquo;s for Pros: it\u0026rsquo;s the only board that puts Linux and display output into a Pico-sized footprint. The hardware design is fully open under CC0, so derivative boards are free to build.\nCons: it\u0026rsquo;s pre-release, so price, purchase channel, and power specs are all unconfirmed. Real-world reliability and performance on PSRAM-backed memory still need verification once hardware ships.\nWho it\u0026rsquo;s for: maker and DIY projects that want a small display or camera preview while keeping the Pico footprint, and embedded developers who want to study an open RISC-V hardware design firsthand. For real-time actuator control on a robot, a microcontroller like the Pico 2 remains the better fit.\nIt\u0026rsquo;s also worth flagging who it is not for, at least yet. Anyone who needs a bill of materials, a firm ship date, or a vendor to buy from today should wait — none of that exists publicly right now. The CC0 release means the schematics and board files can be picked up and cloned by anyone, which is exactly how a lot of small RISC-V boards from Chinese SoC vendors reach a wider audience: not through a single official retailer, but through community fabrication runs once the design settles.\nKorea availability No Korean distribution or launch plans have surfaced — this is an open-hardware project release, not a commercial product announcement yet (TBC). Once it goes on sale, overseas retailers like AliExpress are the most likely first purchase channel.\nVerdict YuzukiNeko is close to the first board to break the \u0026ldquo;Pico-sized means microcontroller\u0026rdquo; assumption. But it\u0026rsquo;s still a pre-release prototype with no confirmed price or power specs — the next step is real hardware proving PSRAM-backed Linux actually runs reliably.\nSources: CNX Software, 2026-09-08, Open Source For You, 2026-09, Raspberry Pi official announcement - Pico 2\n","permalink":"https://dibrain.dev/en/2026/09/09/embedded-yuzukineko-risc-v-pico/","summary":"YuzukiNeko is an open-hardware RISC-V board shaped exactly like a Raspberry Pi Pico, but it runs Linux instead of bare-metal firmware. Here\u0026rsquo;s how it compares to the Pico 2.","title":"YuzukiNeko, Pico's Size With a Linux Brain"},{"content":"Xiaomi officially unveiled its first wide-format book-style foldable, the Xiaomi 18 Fold, in China on September 7. Built around an in-house chip, a Leica-tuned triple camera, and a 6000mAh battery, it\u0026rsquo;s a direct challenge to the big-foldable segment Samsung\u0026rsquo;s Galaxy Z Fold line has held for years.\nBig foldables have improved steadily since the category started, but two numbers barely moved: battery capacity and folded thickness. Samsung\u0026rsquo;s Galaxy Z Fold line has shipped with roughly the same 4400mAh cell since the Z Fold3, and most rivals in the segment have stayed in that same range because the folding hinge and dual-screen stack leave little room for a bigger cell. The Xiaomi 18 Fold is the first phone in the category to break that pattern this decisively, pairing a much larger battery with a body that\u0026rsquo;s actually thinner than most non-folding flagship phones when unfolded.\nXiaomi 18 Fold Spec Detail Cover display 5.38-inch AMOLED, 1712×1168, 120Hz Inner display 7.58-inch flexible AMOLED, 2364×1672, 120Hz Chipset Xring O3 (Xiaomi in-house, 10-core CPU / 16-core GPU) Memory Up to 16GB RAM; the 1TB variant uses LPDDR6 Cameras Leica-tuned, 200MP main + 50MP ultrawide + 50MP periscope telephoto Battery 6000mAh Thickness 5.02mm unfolded / 10.68mm folded Weight 219g Price (China) 16GB+512GB ¥12,999, 16GB+1TB ¥14,999, ceramic 16GB+1TB ¥15,999 Sales start September 10 (China) The base configuration starts at ¥12,999. Whether the phone will reach markets outside China, including Korea, has not been confirmed. The Xring O3 chip is notable on its own: it\u0026rsquo;s Xiaomi\u0026rsquo;s own silicon rather than a licensed Qualcomm or MediaTek part, and the 1TB configuration is among the first phones anywhere to pair it with LPDDR6 memory, a generation ahead of the LPDDR5X still used in most 2026 flagships. Neither the chip\u0026rsquo;s real-world performance nor its thermal behavior inside a folding chassis — where heat has less room to dissipate than in a slab phone — has been independently benchmarked yet.\nRival comparison: Galaxy Z Fold7 Spec Xiaomi 18 Fold Galaxy Z Fold7 Cover display 5.38-inch, 120Hz 6.5-inch FHD+, 120Hz Inner display 7.58-inch, 120Hz 8.0-inch QXGA+, 120Hz Chipset Xring O3 (in-house) Snapdragon 8 Elite for Galaxy Battery 6000mAh 4400mAh Main camera set 200MP + 50MP ultrawide + 50MP tele 200MP + 12MP ultrawide + 10MP tele Thickness (unfolded) 5.02mm TBC Weight 219g TBC Starting price ¥12,999 (China) $1,999 (US) On paper, the Xiaomi 18 Fold leads on battery capacity and camera resolution. But the Galaxy Z Fold7 is already sold in multiple markets with Samsung\u0026rsquo;s water/dust-resistance certification and established after-sales support, while the Xiaomi 18 Fold\u0026rsquo;s launch so far is China-only, with no global distribution or service network confirmed yet.\nThe camera gap is worth a closer look. Both phones use a 200MP main sensor, but the Xiaomi 18 Fold pairs it with a 50MP ultrawide and a 50MP periscope telephoto, versus the Galaxy Z Fold7\u0026rsquo;s 12MP ultrawide and 10MP telephoto. On paper that\u0026rsquo;s a meaningful jump in resolution for the two secondary lenses, though resolution alone doesn\u0026rsquo;t settle image quality — sensor size, lens aperture, and Xiaomi\u0026rsquo;s Leica color tuning versus Samsung\u0026rsquo;s own processing pipeline all matter, and neither camera has been tested side by side yet. The battery gap is more straightforward: 6000mAh against 4400mAh is a roughly 36% increase, which should translate into a noticeably longer day of use, assuming the Xring O3 chip doesn\u0026rsquo;t offset the gain with higher power draw.\nStrengths and weaknesses Strengths\nA 6000mAh battery is a big jump for a big foldable, a segment that has been stuck near 4400mAh. At 5.02mm unfolded, it\u0026rsquo;s slimmer than most phones in its class. The Leica triple-camera setup is unusually capable for a foldable. Weaknesses\nNo confirmed plan for global or Korean availability, so purchase access is uncertain for most readers. Water-resistance rating and hinge-durability certification are absent from the launch materials and remain unverified. Real-world performance and thermals of the new in-house Xring O3 chip haven\u0026rsquo;t been benchmarked independently yet. Who it\u0026rsquo;s for: buyers in China who want the biggest battery and thinnest body currently available in a foldable will find this an easy pick. Buyers elsewhere who value confirmed water resistance, hinge durability, and local after-sales support are still better served by the Galaxy Z Fold7 for now.\nKorea availability No Korean launch plan for the Xiaomi 18 Fold has been confirmed (TBC). Xiaomi\u0026rsquo;s foldable lineup has so far stayed centered on the domestic Chinese market.\nVerdict The Xiaomi 18 Fold tackles two of the foldable category\u0026rsquo;s long-standing weak points — battery size and thickness — in one device, and on spec sheets it beats the Galaxy Z Fold7 in several areas. But global distribution and durability verification are still open questions, so it\u0026rsquo;s worth waiting for an official international launch before treating it as a buying option.\nSources:\nXiaomi 18 Fold Launched: Price, specs and features - SoyaCincau Xiaomi 18 Fold gets its first official look ahead of September 7 launch - GSMArena Xiaomi Just Launched Its First Wide Foldable - Android Headlines Samsung Galaxy Z Fold7 - Samsung US ","permalink":"https://dibrain.dev/en/2026/09/08/phone-xiaomi-18-fold/","summary":"Xiaomi has officially launched its first wide-format foldable, the Xiaomi 18 Fold, in China. A 6000mAh battery and a slimmer hinge put it in direct comparison with Samsung\u0026rsquo;s Galaxy Z Fold7.","title":"Xiaomi 18 Fold: A Thinner Body, a Bigger Battery"},{"content":"Two robot vacuum makers gave two different answers at IFA 2026 in Berlin. Roborock leaned on suction power and a chassis that climbs over thresholds; Ecovacs leaned on keeping the mop pad hygienic while it runs. Both launched in the same week at the same show, which makes them worth comparing spec by spec.\nRoborock Saros 20 Flow Complete / Neo (Qrevo Edge 3 Pro) Suction: 36,000Pa (HyperForce motor) Obstacle avoidance: Reactive AI, recognizes 300+ distinct object types Chassis: AdaptiLift — the robot lifts its own body to clear obstacles and thresholds up to 8.8cm high, and can still squeeze under furniture with as little as 7.95cm of clearance Runtime: up to roughly 350㎡ of coverage per charge (Neo / Qrevo Edge 3 Pro) Price: Qrevo Edge 3 Pro is $1,499.99–$1,500 in the US. Flow Complete\u0026rsquo;s US price is TBC Launch: announced Sept 2–4 at IFA. Flow Complete ships in the US around November 2026 (price not yet announced). In Korea, Flow Complete launches Sept 20 alongside the F25 Pro Turbo Combo and F25 Ultra Steam Gen2 (Korean pricing TBC) Roborock also showed the Saros Rover, a wheel-legged concept robot that attempts to move between floors and clean individual stair steps. It made its European debut at IFA; no price or release date yet. Sources: Ubergizmo, Engadget, Basic-Tutorials The competitor: Ecovacs Deebot X12S OmniCyclone Suction: 27,000Pa, a 23% increase over the previous X12 Battery: 4,000mAh, up to 497 minutes of continuous runtime Mop: a roller-style mop that is continuously rinsed while the robot runs, then self-cleaned at the dock. The dock itself is a bagless OmniCyclone design Price: $1,299.99 Launch: announced Sept 4 at IFA. Korean launch timing is TBC Source: 9to5Toys Neither company had the floor to itself. Dyson used the same week to push its own Nurovi lineup: the R2 Nurovi Wash+Dry launched Sept 3 as a Costco exclusive in the US, with the R3 Nurovi Spot+Scrub UV following on Sept 15. Dyson hasn\u0026rsquo;t published a suction figure that lines up cleanly with Roborock\u0026rsquo;s or Ecovacs\u0026rsquo; Pa ratings, so it\u0026rsquo;s mentioned here as market context rather than a third contender in the spec comparison below. (Source: VacuumWars)\nHead to head Spec Roborock Qrevo Edge 3 Pro Ecovacs X12S OmniCyclone Suction 36,000Pa 27,000Pa Signature feature AdaptiLift chassis (clears 8.8cm thresholds) Continuously-rinsed roller mop Runtime ~350㎡ per charge Up to 497 minutes US price $1,499.99–$1,500 $1,299.99 Korea launch Sept 20 (Flow Complete; price TBC) TBC Trade-offs and who each one fits Roborock\u0026rsquo;s number is bigger on paper, and the AdaptiLift chassis is specifically built to get over thresholds and rug edges without help — a real advantage in Korean apartments, which tend to have more raised thresholds between rooms than typical US or European floor plans. The catch is a roughly $200 higher price and no confirmed Korean price yet, so a final verdict has to wait.\nEcovacs trails on raw suction, but a mop that gets rinsed continuously while it\u0026rsquo;s still cleaning is a concrete, practical difference for households that mop often — especially ones with pets, where a dirty mop pad being dragged around the floor is the actual complaint. It\u0026rsquo;s also $200 cheaper.\nBoth robots use lidar-based obstacle avoidance, but Roborock is the only one that put a number on it (300+ recognized object types); Ecovacs hasn\u0026rsquo;t published an equivalent figure for this launch, so a direct comparison on avoidance accuracy isn\u0026rsquo;t possible yet.\nThere\u0026rsquo;s also a naming wrinkle worth flagging before anyone shops by model name: the same underlying hardware that Roborock sells internationally as the Saros 20 Neo is badged Qrevo Edge 3 Pro in the US, while the Flow Complete is a separate, higher-tier model in the same family that\u0026rsquo;s getting the Korean launch. If you\u0026rsquo;re comparing prices across regions, check the exact model number, not just the \u0026ldquo;Saros 20\u0026rdquo; family name.\nAvailability in Korea Roborock\u0026rsquo;s Saros 20 Flow Complete has a confirmed Korean launch date of Sept 20, alongside the F25 Pro Turbo Combo and F25 Ultra Steam Gen2, though Korean pricing hasn\u0026rsquo;t been announced. Ecovacs has not confirmed a Korean launch date or price for the Deebot X12S OmniCyclone.\nVerdict At equal pricing, Roborock\u0026rsquo;s suction and threshold-clearing chassis make it the better fit for Korean homes. But with a $200 gap and no Korean price yet, households that prioritize mop hygiene have a reasonable case for waiting on Ecovacs\u0026rsquo; Korean launch details instead. The stair-climbing Saros Rover is still a concept and didn\u0026rsquo;t factor into this comparison, but it\u0026rsquo;s worth watching as a sign of where robot vacuums go next.\n","permalink":"https://dibrain.dev/en/2026/09/07/appliance-roborock-saros20-ecovacs-x12s/","summary":"Roborock\u0026rsquo;s Saros 20 lineup and Ecovacs\u0026rsquo; Deebot X12S OmniCyclone both launched at IFA 2026. We compare suction power, obstacle-clearing chassis design, mop hygiene, and price.","title":"Roborock Pulls Harder, Ecovacs Rinses Cleaner"},{"content":"On the last day of the first week of September, Samsung\u0026rsquo;s Galaxy S26 FE officially went on sale in Korea. Abroad, a brainwave-reading wearable called Atlas 1.0 already started shipping, and in software, Upstage\u0026rsquo;s large open model Solar Open2-250B was the talk of the week. None of the three is a dramatic leap on its own, but together they sketch where each part of the market is spending its effort right now: Samsung repackaging known hardware at a lower price, a small hardware maker pushing into an unproven sensor category, and Korean AI labs racing to field their own large open models alongside global names. Links to this week\u0026rsquo;s earlier deep dives on an embedded board, a GitHub pick, and a wearable are collected at the bottom.\nKorea Samsung Galaxy S26 FE Chipset: Exynos 2500 Memory/storage: 8GB RAM, 128GB or 256GB Display: 6.7-inch AMOLED, 2340×1080, 120Hz Cameras: 50MP main + 8MP ultrawide + 8MP 3x telephoto (rear), 12MP (front) Battery: 4,900mAh Build: 161.6×76.9×7.4mm, aluminum frame, IP68 Colors: Blueberry, Graphite, Pistachio Price: KRW 1,045,000 for 256GB (128GB price TBC) Release: unveiled August 27, on sale in Korea from September 4 through Samsung Store, Samsung.com, and carrier stores Take: an Exynos 2500 and 8GB of RAM is the same combination the mainline Galaxy S series already shipped last year, so the FE is doing exactly what an FE is supposed to do — selling last year\u0026rsquo;s flagship at today\u0026rsquo;s price. Keeping IP68 and an aluminum frame while dropping into the low-1-million-won range is the actual news here; this is a phone for buyers prioritizing durability and price over the newest camera or AI features. The triple rear camera setup (50MP main, 8MP ultrawide, 8MP 3x telephoto) also matches last year\u0026rsquo;s tier rather than this year\u0026rsquo;s flagship, which is consistent with how Samsung has positioned every FE model since the line started: recent flagship silicon and design, but camera hardware and RAM held back a generation to hit a lower price point. Sources: 9to5Google, GSMArena, Edaily, Gukje News Global Atlas 1.0 (Brain Wearable) Category: EEG-based wearable Price: $499 Release: went on sale September 1 (local time) Sensor channel count, battery life, companion app, and data-processing approach: TBC Take: consumer EEG wearables are still an unproven category — accuracy claims and real use cases haven\u0026rsquo;t been independently verified yet. A $499 price tag alone doesn\u0026rsquo;t tell you whether this is a genuine multi-channel EEG device or a simplified focus-tracking gadget, so the specs are worth treating with caution until hands-on reviews and clinical backing show up. The broader pattern worth watching is that \u0026ldquo;brain wearables\u0026rdquo; keep launching at consumer price points well before the underlying signal-processing claims get third-party validation — the same caution that applied to earlier sleep- and focus-tracking headbands applies here. Source: Gear Patrol Software Solar Open2-250B (Upstage) Developer: Upstage (South Korea) Scale: 250B parameters (per the model name) Purpose: high-performance enterprise reasoning, open deployment License, exact release date, and benchmark numbers: TBC On-device: a 250B-class model can\u0026rsquo;t run on a laptop or edge device even in a quantized form — it assumes server or cloud deployment with multiple accelerators, which puts it in a different bracket from the smaller on-device models this blog usually covers on Thursdays Take: this week saw a wave of large open models land at once. Z.ai\u0026rsquo;s GLM-5.2 is drawing attention as a mixture-of-experts model tuned for coding, math, and agent tasks; DeepSeek-V4-Flash is being adopted for low-latency assistants and coding tools; and Moonshot AI\u0026rsquo;s Kimi-K3 is positioned for long-context, document-heavy RAG pipelines. Among that group, Upstage\u0026rsquo;s entry is the one most likely to show up in Korean enterprise software stacks, given the company\u0026rsquo;s existing customer base here. Adoption decisions should still wait until the license and real benchmark numbers are confirmed — a parameter count in a model\u0026rsquo;s name is not a substitute for an independent evaluation. Source: Tech AI Magazine This Week at a Glance Product Category Region Key specs Price Release Galaxy S26 FE Smartphone Korea Exynos 2500, 6.7\u0026quot;, 4,900mAh KRW 1,045,000 (256GB) Sep 4 Atlas 1.0 AI wearable Global EEG-based sensor $499 Sep 1 Solar Open2-250B Open-source LLM Software 250B parameters Open model TBC This Week on DigitalBrain Jetson Orin Nano 2: the 2x that isn\u0026rsquo;t MiniMind Thinks Small, God\u0026rsquo;s Eye Sees Big Huawei Watch 6: Brighter Screen, Narrower World ","permalink":"https://dibrain.dev/en/2026/09/04/brief-galaxy-s26-fe-atlas/","summary":"Samsung\u0026rsquo;s Galaxy S26 FE went on sale in Korea at KRW 1,045,000, a brainwave-reading wearable called Atlas 1.0 launched abroad, and Upstage\u0026rsquo;s large open model Solar Open2-250B drew attention in software.","title":"Galaxy S26 FE Costs Less, a Wearable Reads More"},{"content":"Two repositories climbed GitHub\u0026rsquo;s trending chart this week heading in opposite directions. One shrinks a language model down to something you can train on a single desktop GPU; the other pulls the planet\u0026rsquo;s live data feeds into one browser tab. Here\u0026rsquo;s a closer look at both, picked with on-device AI and developer tooling in mind — the kind of projects worth cloning on a weekend rather than just reading about.\nMiniMind Repo: jingyaogong/minimind License: Apache License 2.0 Language: Python Stars: 57,757 (at time of writing), +1,949 this week Latest release: minimind-3 / minimind-3-moe (2026-04-01) MiniMind trains a 64-million-parameter language model from scratch using plain, unwrapped code instead of a heavyweight framework. The README pegs the cost at roughly 3 yuan (about $0.43) and two hours on a single RTX 3090 (24GB). Both the training and inference scripts are short enough to read end to end, which makes this a project for developers who actually want to trace what happens inside a transformer training loop rather than call an API.\ngit clone --depth 1 https://github.com/jingyaogong/minimind cd minimind \u0026amp;\u0026amp; pip install -r requirements.txt # Inference python eval_llm.py --load_from ./minimind-3 # Training cd trainer \u0026amp;\u0026amp; python train_pretrain.py cd trainer \u0026amp;\u0026amp; python train_full_sft.py On C-Eval and CMMLU the base model scores 24.89% and 25.38% respectively; on English benchmarks (ARC, PIQA, OpenBookQA, HellaSwag, SocialIQA) scores range from roughly 23% to 50% depending on the task, per the README. A LoRA-aligned variant, minimind-3-exam, adds about 2.9 percentage points on top of that. At 64M parameters this isn\u0026rsquo;t production-grade output — it\u0026rsquo;s a teaching tool, best suited to students and engineers who want to build LLM training intuition line by line rather than deploy something.\nWhat makes this worth a look isn\u0026rsquo;t the benchmark numbers, which are modest by design, but the fact that the entire stack — tokenizer, pretraining loop, supervised fine-tuning, and evaluation — fits in a repository small enough to read in an afternoon. Most public LLM training code lives inside large frameworks that hide the actual mechanics behind configuration files and abstraction layers. MiniMind strips that away, so a single RTX 3090 and a couple of hours are enough to watch loss curves move and see exactly which line of code caused it. That tradeoff — clarity over capability — is the whole pitch.\nGod\u0026rsquo;s Eye View Repo: bilawalsidhu/gods-eye-view License: MIT Language: JavaScript Stars: 16,070, +12,042 this week (the single biggest weekly gain on the trending page) Latest release: TBC — the repo ships by commit rather than tagged releases God\u0026rsquo;s Eye View is a browser-based spatial-intelligence dashboard that layers public real-time data onto a photorealistic 3D globe: aircraft from OpenSky Network and adsb.lol, vessels from AISStream, satellites from CelesTrak, earthquakes from USGS, wildfires from NASA FIRMS, launch schedules from Launch Library 2, city CCTV feeds, and geolocated radio stations. Basemaps come from Cesium ion, Google Maps, Esri, and OpenStreetMap.\nnpm install npm run doctor npm run dev # open http://localhost:4173 It can also be installed through Pinokio by pasting the repo URL into the Discover tab. This is worth a look for frontend developers who want to build a live tracking dashboard without standing up a dedicated GIS stack, or anyone looking for a side project built entirely on public data APIs. As a pure open-source tool, there\u0026rsquo;s no Korea-specific launch or pricing angle to report here.\nThe interesting engineering problem here isn\u0026rsquo;t the 3D rendering — Cesium already handles that — it\u0026rsquo;s normalizing a dozen unrelated data sources with different rate limits, coordinate systems, and update frequencies into one coherent scene. A satellite feed from CelesTrak refreshes on a very different cadence than an OpenSky aircraft ping, and the repo\u0026rsquo;s real contribution is the plumbing that keeps all of it in sync without the browser choking. That\u0026rsquo;s also the biggest weekly star gain on this week\u0026rsquo;s entire trending page, ahead of every other repository — a sign that \u0026ldquo;open data, one screen\u0026rdquo; is resonating well beyond the GIS niche it started in.\nWrap-up These two projects pull in opposite directions but teach a similar lesson. MiniMind shrinks the model so you can actually see the training process; God\u0026rsquo;s Eye View stitches together a dozen public APIs so you can see what\u0026rsquo;s possible when you compose data instead of building it from scratch. Both repos run as-is — clone whichever one matches what you\u0026rsquo;re curious about and try it.\n","permalink":"https://dibrain.dev/en/2026/09/03/dev-minimind-gods-eye-view/","summary":"This week\u0026rsquo;s GitHub trending list pairs MiniMind, a 64M-parameter language model you can train in two hours, with God\u0026rsquo;s Eye View, a browser-based 3D globe streaming live satellite, flight, and shipping data.","title":"MiniMind Thinks Small, God's Eye Sees Big"},{"content":"Huawei used IFA 2026\u0026rsquo;s media days (September 2-3) to unveil the Watch 6 and Watch 6 Pro, and much of the coverage frames it as a direct answer to Google\u0026rsquo;s Pixel Watch 5, launched three weeks earlier. The same event brought a crowded wearables lineup: TECNO showed its Watch 4 and Buds 5, Huawei also refreshed the Watch GT 7, and Samsung, Withings, RingConn and Ultrahuman are all expected on the show floor. Amid that crowd, the Watch 6 stands out on brightness and case materials, but the numbers that actually decide day-to-day usability — battery life, water resistance — weren\u0026rsquo;t part of this week\u0026rsquo;s announcement.\nSource: Gadgets \u0026amp; Wearables Huawei Watch 6 / Watch 6 Pro Spec Watch 6 Watch 6 Pro Case 42mm / 46mm Ceramic or titanium Weight / thickness 38g, 10.2mm (lightest configuration) TBC Display OLED (brightness TBC) OLED, up to 3,500 nits Sensors Standard health sensor suite ECG, upgraded X-Tap sensor Battery TBC TBC Water resistance TBC TBC OS / compatibility Huawei\u0026rsquo;s own OS stack, works with iOS and Android (version TBC) Same Price (UK) £399.99 £599.99 Availability UK sales from September 23; £50 off if purchased Sep 23–Oct 26 Same Huawei\u0026rsquo;s own weight and thickness figures aren\u0026rsquo;t tied to a specific case size in the launch material, so we\u0026rsquo;re labeling them \u0026ldquo;lightest configuration\u0026rdquo; rather than assigning them to one model. Battery capacity and water-resistance rating were not disclosed in this week\u0026rsquo;s announcement, which is notable because those two numbers are usually the first thing Huawei leads with on its GT line — the GT 6 series, launched in Korea back in May, was marketed heavily around up to 21 days of battery life. Their absence here suggests Huawei may still be finalizing those figures, or is saving them for a later regional rollout.\nThe Pro model\u0026rsquo;s headline sensor is described as an \u0026ldquo;improved X-Tap sensor,\u0026rdquo; which implies a previous version exists in Huawei\u0026rsquo;s lineup, but exactly what changed — which metrics it adds or how accuracy improved — wasn\u0026rsquo;t detailed in the sources available this week. Until Huawei publishes a fuller spec sheet, treat that claim as directional rather than measured.\nSources: TechAdvisor, Gizmochina, Android Headlines Vs. Pixel Watch 5 Spec Huawei Watch 6 Google Pixel Watch 5 Announced September 2, 2026 (IFA) August 12, 2026 (Made by Google) Case sizes 42mm / 46mm 41mm / 45mm Storage TBC 64GB (double the previous generation) Coprocessor TBC Cortex-M55 Health features ECG (Pro), X-Tap sensor Blood pressure trends, breathing emergency detection, cloud-assisted GPS (Google claims 2x accuracy) OS ecosystem Huawei\u0026rsquo;s own OS, no Google Mobile Services Wear OS, full Google service integration Price £399.99 / £599.99 $399–$529 (41/45mm, WiFi/LTE), Special Edition $579 Both watches arrived as their maker\u0026rsquo;s flagship reveal for the season — Huawei\u0026rsquo;s at IFA, Google\u0026rsquo;s at its own hardware event three weeks prior. The real divide isn\u0026rsquo;t the spec sheet, though: Pixel Watch 5 runs Wear OS with full access to Google\u0026rsquo;s app ecosystem, while Watch 6 runs on Huawei\u0026rsquo;s own OS without Google Mobile Services. An Android phone owner expecting Google Maps or Assistant on their wrist should not assume Watch 6 will deliver that, and an iPhone owner should check compatibility details carefully before assuming full feature parity with Apple\u0026rsquo;s own watch software.\nSources: Droid-Life, GSMArena Trade-offs and who it\u0026rsquo;s for Strengths: The Watch 6 Pro\u0026rsquo;s 3,500-nit display is among the brightest in this price bracket, which should matter a lot for outdoor legibility in direct sunlight. The ceramic/titanium case options look more premium than most competitors at this price, and the ECG/X-Tap sensor pairing gives the Pro model a health-tracking angle beyond what a standard optical heart-rate sensor offers. Weaknesses: No Google Mobile Services makes it a harder sell for anyone who relies daily on Google Maps, Assistant, or Calendar on their wrist. Key numbers — battery life, water resistance — aren\u0026rsquo;t public yet, so real-world stamina during workouts or swimming is still unverified. Best for: People already using a Huawei phone or Huawei\u0026rsquo;s app ecosystem, or anyone who prioritizes screen brightness and case materials over ecosystem convenience. If you\u0026rsquo;re Google-service-dependent on Android, the Pixel Watch 5 remains the safer, better-integrated pick. Regional availability There was no confirmation of Korean availability for the Watch 6 or Watch 6 Pro in this week\u0026rsquo;s launch material. Huawei\u0026rsquo;s Korean unit launched the separate, golf-focused Watch GT 6 series domestically back in May, so the local distribution channel exists, but the Watch 6 is a distinct mainstream lineup from the GT line — its Korean launch, pricing, and timing all remain TBC. Converting the UK price (£399.99 / £599.99) into won to guess a local price would be speculation; waiting for an official announcement is the safer read.\nSources: Digital Chosun, Asia Economy Verdict The Huawei Watch 6 leads with a bright display and premium materials, but it hasn\u0026rsquo;t shown its hand on the basics — battery life and water resistance — yet. Given how many rivals launched in the same week (Pixel Watch 5, TECNO Watch 4, and more still to come at IFA), this announcement looks like the opening move of the season\u0026rsquo;s spec race, not the final word. If you\u0026rsquo;re already inside Google\u0026rsquo;s ecosystem, there\u0026rsquo;s little reason to switch. If you\u0026rsquo;re a Huawei app user, the Pro model\u0026rsquo;s display alone is a real reason to consider it. We\u0026rsquo;ll revisit this once the missing numbers — and a Korean launch decision — are public.\n","permalink":"https://dibrain.dev/en/2026/09/03/wearable-huawei-watch-6/","summary":"Huawei\u0026rsquo;s IFA 2026 launch of the Watch 6 and Watch 6 Pro, compared against Google\u0026rsquo;s Pixel Watch 5 on specs, price, and ecosystem lock-in.","title":"Huawei Watch 6: Brighter Screen, Narrower World"},{"content":"NVIDIA announced the Jetson Orin Nano 2 on August 25, with the module and developer kit expected in the first half of 2027. The headline is \u0026ldquo;2x inference performance,\u0026rdquo; which reads like a generational jump. Line up the specs against the board it replaces and the picture is more specific than that, and the numbers that matter for a robot build are not the ones in the headline.\nSide by side Orin Nano 2 (new) Orin Nano Super (current) AI compute 78 TOPS 67 TOPS GPU Ampere, 1,536 CUDA cores Ampere, 1,024 CUDA cores + 32 tensor cores CPU 8-core Arm Cortex-A78 6-core Arm Cortex-A78AE Memory 8 GB LPDDR5X, 120 GB/s 8 GB 128-bit LPDDR5, 102 GB/s Power 15 W to 40 W 7 W to 25 W Price TBC (not announced) $249 developer kit Availability H1 2027 Shipping now Sources: NVIDIA Newsroom, Hackster.io, NVIDIA Jetson Orin Nano Super Developer Kit\nThe \u0026ldquo;2x\u0026rdquo; is not measured at equal power Start here, because it changes how you read everything else. On paper the TOPS figure moves from 67 to 78, about 16%. The doubling claim comes from a different comparison: the new board in its 40 W mode against the old board at its 25 W ceiling. It is not twice the performance at the same power draw.\nNVIDIA\u0026rsquo;s other claim is the more useful one: 40% lower power at equivalent performance. Put together, the new module gives you more when you feed it more, and costs less when you ask for the same. That also tells you what did not happen. The GPU is still Ampere. The gains come from 50% more CUDA cores, two extra CPU cores, and 18% more memory bandwidth, not from a new architecture.\nThe power floor is the part that bites This is the quietest row in the table and the one most likely to break a design. The envelope moves from 7-25 W to 15-40 W. The ceiling going up is expected. The floor doubling, from 7 W to 15 W, is the problem.\nIf you have built a battery-powered mobile robot, you already know why. Idle and low-load draw sets your runtime as much as peak draw does. A design that keeps a camera alive and only runs inference on an event spends most of its life near the floor. Swapping in the new module doubles that baseline, and the battery budget you sized for the old board no longer holds.\nThe 40 W ceiling deserves the same caution from the other direction. You cannot drop this module into a power section designed around 25 W and call it done: rail capacity, connector ratings, and thermal design all need to be recalculated. The module\u0026rsquo;s input voltage and connector pinout have not been published yet, so treat those as TBC until the datasheet lands. Guessing at a power section is how boards get cooked.\nMemory stays at 8 GB Both generations ship 8 GB. Bandwidth improves from 102 to 120 GB/s, but capacity does not move. Run vision models on-device for a while and you find the wall is usually memory, not compute. Jetson shares memory between CPU and GPU, so after the OS and the camera pipeline take their share, what is actually available to a model is well under 8 GB.\nIf the plan is to quantize a current vision-language model and run it at the edge, this generation hits the same wall as the last one. Memory fills before you get anywhere near 78 TOPS. More capacity means moving up to something like Orin NX, which is a different price bracket and a different conversation.\nWhat to buy right now Shipping is H1 2027. That is more than six months out, with no announced price. If you have to ship something in 2026, this board is not on the table.\nBuilding now: the Orin Nano Super developer kit is $249, it is not short on performance, and you can buy it today. Its 7 W floor is also friendlier to battery-powered work than the new board\u0026rsquo;s 15 W. Targeting production in 2027 or later: plan around the new module, but design the power section and cooling for 40 W from the start. Adding headroom later costs far more than building it in. Learning or teaching: no reason to wait. The 8 GB ceiling is the same on both, and what you learn transfers when the board changes. Verdict Honestly read, this is a widened Ampere part rather than a new generation: more cores, a broader power envelope, faster memory. That is a perfectly good product. The risk is taking the \u0026ldquo;2x\u0026rdquo; line at face value and carrying it into a design, because the place it will hurt you is power. Work out whether your enclosure can actually deliver 40 W, and whether your battery can live with a 15 W floor, before anything else.\nWorth revisiting once pricing and the full module datasheet are published.\n","permalink":"https://dibrain.dev/en/2026/09/02/embedded-jetson-orin-nano-2/","summary":"NVIDIA\u0026rsquo;s next entry-level Jetson compared against the current Orin Nano Super, with a look at the power envelope and memory ceiling that decide whether it fits a robot.","title":"Jetson Orin Nano 2: the 2x that isn't"},{"content":"The first week of September sits in the calm before two big dates: Apple\u0026rsquo;s September 9 event and IFA Berlin, which opens on September 4. Even so, four products stood out this week. In Korea, LG and Samsung launched budget 14-inch laptops on the same day, which tells you where the domestic PC fight has moved. Overseas, NVIDIA refreshed its entry-level Jetson for robotics, and Poco launched the F9 series with a price increase that is hard to ignore.\nKorea LG gram Book AI 2026 14 (14U40V) LG\u0026rsquo;s gram line has always been about weight, and the gram Book sub-brand is its cheaper, thicker sibling. This is the first time the gram Book line gets a 14-inch model.\nCPU: Intel Core Series 3 (Wildcat Lake) Battery: 61.2 Wh, up to 30.5 hours by LG\u0026rsquo;s own measurement Weight / thickness: 1.29 kg / 14.9 mm, aluminum chassis Display: 14-inch. Resolution and panel type: TBC Memory / storage: TBC (varies by configuration) Price: from 1,450,000 KRW (MSRP) Release: announced Sep 1; on sale Sep 7 at LGE.com and major online malls, with LG Best Shop stores following from mid-September Take: 1.29 kg with a 61 Wh pack is a strong ratio for a student machine, and Wildcat Lake should help idle power. But LG has not yet published the display resolution or RAM tiers, so I would hold off on judging the price until the full spec sheet is out. Source: The Korea Economic Daily, THE ELEC Samsung New Galaxy Book6 Samsung\u0026rsquo;s answer landed the same morning. The \u0026ldquo;New\u0026rdquo; Galaxy Book6 is a single-size, single-color budget model aimed at students heading back to school.\nCPU: Intel Core Series 3 Display: 14-inch (35.6 cm), one size, Gray only Weight: 1.35 kg Battery: up to 25 hours of video playback; a 30-minute charge restores up to 33% Price: 1,190,000 to 1,490,000 KRW depending on CPU and OS choice Release: Sep 1 in Korea, through Samsung Stores, Samsung.com, and open-market retailers Take: Two Korean giants launching on the same day, in almost the same price band, means the mid-range market is now the main battleground at home. The Galaxy Book6 is 60 grams heavier than the gram Book, but its lower entry price will matter more to first-time buyers than the weight difference. Source: Money Today, Edaily Global NVIDIA Jetson Orin Nano 2 NVIDIA\u0026rsquo;s entry-level Jetson has been the default brain for hobby and education robots since the Orin Nano Super price cut. The Orin Nano 2 keeps the same tier but changes the silicon underneath.\nAI compute: 78 TOPS CPU: 8-core Arm Cortex-A78 GPU: Ampere architecture, 1,536 CUDA cores Memory: 8 GB LPDDR5X at 120 GB/s Power: 15 W to 40 W modes NVIDIA claims 2x inference performance over the previous generation, and 40% lower power at the same performance level Price: TBC (not announced) Release: announced Aug 25 (US time); the module and developer kit are expected in the first half of 2027 Take: The GPU is still Ampere, so the gain comes from more CUDA cores, two extra CPU cores, and faster memory rather than a new architecture. For an on-device vision robot, 8 GB of shared memory is once again the real ceiling. The new 40 W top mode is the part that affects hardware design: a small robot built around the old 25 W budget needs its power rail and cooling rethought before this module goes in. And with shipping a good six months away, nobody should plan a 2026 build around it. Source: NVIDIA Newsroom, Hackster.io Going deeper: SBC Deep Dive: What Actually Changed in the Jetson Orin Nano 2 covers the comparison and the power-design side. Poco F9 Pro and F9 Ultra Poco built its name on flagship chips at mid-range prices. The F9 series keeps the flagship chip but moves the price up a tier.\nSoC: Snapdragon 8 Elite Gen 5 on both models Display: 6.59-inch (Pro) / 6.9-inch (Ultra), 185 Hz, 4,500 nits peak, Gorilla Glass 7i Camera: 200 MP main with OIS, 32 MP front Battery: 6,330 mAh (Pro) / 8,050 mAh (Ultra), 100 W wired and 50 W wireless charging IP68 rating, Bose-tuned speakers Price: F9 Pro from $699 (12 GB / 256 GB), F9 Ultra from $799; UK £799 / £899. Euro pricing differs between sources (€799 / €999 versus €899.90 / €1,099.90): TBC. India pricing: TBC Release: global launch Sep 1 Take: An 8,050 mAh cell with 100 W charging in a phone body is the headline, and the 185 Hz panel is a spec nobody asked for but gamers will like. The problem is price. Poco itself blamed rising memory costs, but with the same chip as several rivals, the value story that defined the F series is much weaker this year. Source: 9to5Google, Tech Advisor Software Pollen Robotics microduck Pollen Robotics, now part of Hugging Face, published the software stack for microduck, a 25 cm, 800 g biped duck robot that walks, rolls, and grasps using reinforcement-learning policies. The repository trended on GitHub this week after the August 28 announcement.\nRepo: github.com/pollen-robotics/microduck, Apache-2.0, Rust, about 6.3k stars at the time of writing Runs on a Rockchip RK3566 board with a 50 Hz control loop driving 15 servos, plus camera, ToF depth sensor, Bluetooth, and gamepad support Companion repo microduck_rl holds the simulation and RL training environments Install or run: the README points to the documentation and a robotctl command-line tool rather than a one-line install; the robot itself is sold at pollen-robotics.com/microduck Take: The interesting part for education is the pipeline, not the duck: train a gait in simulation, deploy to a cheap ARM board, iterate. Note that only the software is open; Pollen has asked press not to call the hardware open source, so do not expect STL files. Source: GitHub, CNX Software This Week at a Glance Product Category Region Key specs Price Release LG gram Book AI 2026 14 Laptop Korea Core Series 3, 61.2 Wh, 1.29 kg from 1,450,000 KRW Sep 7 Samsung New Galaxy Book6 Laptop Korea Core Series 3, 14-inch, 1.35 kg 1,190,000 to 1,490,000 KRW Sep 1 NVIDIA Jetson Orin Nano 2 SBC / Robotics Global 78 TOPS, 8-core A78, 8 GB LPDDR5X TBC H1 2027 Poco F9 Pro / F9 Ultra Smartphone Global Snapdragon 8 Elite Gen 5, 185 Hz, 6,330 / 8,050 mAh from $699 / $799 Sep 1 Pollen Robotics microduck Open source Software Rust, RK3566, 15 servos, RL policies Apache-2.0 Aug 28 ","permalink":"https://dibrain.dev/en/2026/09/02/brief-weekly-gadgets/","summary":"Two budget 14-inch laptops launched in Korea on the same day, NVIDIA\u0026rsquo;s next entry-level Jetson for robots, and Poco\u0026rsquo;s pricier F9 flagships.","title":"On sale today, shipping next year"},{"content":"DigitalBrain DigitalBrain covers newly released hardware and software with the specs, prices, and release dates in one place, each entry sourced. One post every weekday, with a fixed subject per day.\nDay Subject Monday Laptops, phones, appliances or wearables — whatever launched Tuesday The same range, a different category from Monday Wednesday Single-board computers, dev boards, robots Thursday Dev picks (trending GitHub repositories / Hugging Face models) Friday The week in brief (Korea / Global / Software) The rules this site holds to:\nOnly products and projects that were actually announced are covered, and every entry links to its source. Specs follow the manufacturer\u0026rsquo;s official figures. Anything unconfirmed is marked \u0026ldquo;TBC\u0026rdquo; rather than filled in. When there is nothing genuinely new that day, nothing is published. No filler to keep a streak alive. Short opinions are the author\u0026rsquo;s own and are not purchase recommendations. The Korean and English posts are written separately from the same research, each for its own readers. Who writes this I write here as yjworks. I work with embedded software and computer vision, and I have built boards and robots myself. That background is why these posts care less about the spec sheet and more about whether a thing actually behaves the way it is advertised.\nThis is a personal blog. It is not affiliated with any company or manufacturer, and every opinion here is my own.\nContact Email: leeyunjai1982@gmail.com GitHub: github.com/yjworks Corrections and tips: email me or open an issue. If a spec or price is wrong, send the source and I will fix it. ","permalink":"https://dibrain.dev/en/about/","summary":"About the DigitalBrain blog","title":"About"},{"content":"DigitalBrain (\u0026ldquo;this site\u0026rdquo;, https://dibrain.dev/) respects your privacy. This policy explains what information is collected when you visit this site, how it is used, and the choices you have.\n1. Information We Collect This site has no user accounts or login, and does not ask visitors for personal information such as names or email addresses.\nHowever, the third-party services described below may automatically collect the following information through cookies and similar technologies:\nIP address (which may be anonymized), browser type and version, operating system Date and time of visit, pages viewed, referring URL Advertising identifiers and cookie identifiers 2. Cookies Cookies are small text files that a website stores in your browser. This site uses cookies for the following purposes:\nAnalyzing site traffic and usage Serving advertisements relevant to your interests and measuring ad performance You can refuse or delete cookies in your browser settings. Refusing cookies does not limit your ability to read this site.\n3. Google AdSense This site uses Google AdSense, an advertising service provided by Google.\nThird-party vendors, including Google, use cookies to serve ads based on your prior visits to this site or other websites. Google\u0026rsquo;s use of advertising cookies enables it and its partners to serve ads based on your visits to this site and/or other sites on the Internet. You may opt out of personalized advertising by visiting Google Ads Settings. You may opt out of third-party vendors\u0026rsquo; use of cookies for personalized advertising at www.aboutads.info. For details on how Google uses data, see How Google uses information from sites or apps that use our services.\n4. Google Analytics This site may use Google Analytics to analyze visitor statistics.\nGoogle Analytics uses cookies to collect information about how visitors use this site. That information is transmitted to and stored on Google\u0026rsquo;s servers. The collected information is used only to understand site usage and improve the content. You can opt out by installing the Google Analytics Opt-out Browser Add-on. 5. How Information Is Used Collected information is used only for:\nAnalyzing site usage and improving content Serving advertisements and measuring their performance Preventing abuse and maintaining security 6. Third Parties The site operator does not directly collect or store visitor information, and does not share information with any third party other than the Google services described above. Google\u0026rsquo;s handling of the information it collects is governed by the Google Privacy Policy.\n7. External Links Posts on this site contain links to external sites such as manufacturer pages, news articles, and open source repositories. Those sites have their own privacy practices, for which this site is not responsible.\n8. Children\u0026rsquo;s Privacy This site is not directed at children under 14 and does not knowingly collect personal information from children.\n9. Changes to This Policy This policy may be updated to reflect changes in law or services. Any changes will be posted on this page.\n10. Contact For privacy-related questions, please contact:\nOperator: yjworks\nEmail: leeyunjai1982@gmail.com\nGitHub: yjworks/yjworks.github.io\nEffective date: September 2, 2026\n","permalink":"https://dibrain.dev/en/privacy/","summary":"Privacy policy for the DigitalBrain blog","title":"Privacy Policy"}]