VOL.2026-09-10 · 38 STORIES · SKILLHUB DAILY
SkillHub.news Daily AI
Sep 10, 2026 · #2 · updated 08:00 daily · sources 5
01MODEL RELEASES6 stories
DeepSeek V4.1 Flash goes GA, with simultaneous price cuts across the Flash line.
DeepSeek官方Confidence HighProgress · Day 2
The limited-time beta endpoint goes GA as scheduled, outperforming V4 Pro across the board; during the transition, V4 Pro requests are automatically routed to Flash and billed at Flash rates.
Take · Price-cut structure correction: cache hits -60% (0.05→0.02 yuan/million), cache misses -33%, output only -11% (4.5→4 yuan); 'output down 60%' was misinformation—the intent is to steer high-frequency Agent workflows to the cache-hit path.
Ant open-sources Bailing Ling-3.0-flash-VL: native multimodal, only 5.5B active.
蚂蚁百灵官方+多源Confidence High
MoE 124B total parameters / 5.5B active, native image, text, and video input, 256K context, first 'observe-act-verify-correct' visual feedback loop, surpasses GPT-5.4 on Image-to-WebDev Arena.
Take · Small-activation multimodal lowers the barrier to low-cost inference another notch, with a single-GPU-runnable multimodal alternative emerging.
Cohere open-sources megakernel inference engine: 1.58× faster than vLLM.
Cohere官方博客Confidence MidSingle-source pending verification
decode-fused single kernel, measured at 292 tok/s on H100 BF16 batch1, Apache 2.0 license.
Take · Inference infrastructure is becoming one of the high grounds as ecosystem value migrates from the 'weights layer' to the 'toolchain layer'.
Tencent open-sources TeamAI CLI: centralized management of team-level skills / MCP / hooks
腾讯开源社区Confidence MidSingle-source pending verification
Aimed at unified team management of Agent skills, MCP, and hooks, the repo gained 425 stars in 24 hours.
Take · Team-level infrastructure is beginning to emerge in the skill distribution layer, creating a division of labor with the individual skill market.
Anthropic reportedly launches MHS hardware standard
Anthropic外媒Confidence MidSingle-source pending verification
Foreign media report Anthropic is involved in advancing a model-oriented hardware standard (MHS); details not confirmed by a second source.
OpenAI discloses single-project cost of a 10,000-level agent math push
OpenAI财联社Confidence Mid
Approx. 2.7 million messages / 130 billion token / approx. $10 million (day 2 increment of the Navier-Stokes push).
Take · The 70,000 messages and the $400,000 investigation related to METR are two separate events; don't conflate them; leading labs begin publicly accounting per project, and the market enters an ROI calculation phase.
02PRODUCT4 stories
JD.com JDD physical AI triple play: open-source world model + 100,000-card AI compute + million-scale robot procurement
京东官方+多源Confidence High
Open-source world model JoyAI-Echo WM (No. 1 in WBench navigation with 81.6 points); planning with Moore Threads to co-build a 100,000-card domestic AI compute cluster; 5-year logistics procurement of 3 million robots / 1 million unmanned vehicles / 100,000 drones.
Take · The AI race is shifting from the model layer to the physical world and heavy-asset order layer—a rare simultaneous three-front push by a domestic vendor across 'model + compute + orders.'
XPeng IRON mass-production validation completed: first humanoid robot automated production line runs successfully
小鹏小鹏官方Confidence HighProgress · Day 2
World's first humanoid robot automated production line runs successfully; first unit rolls off the line after independent final assembly (automation rate over 80%); mass production by year-end; 2027 overseas delivery commitment unchanged; automotive-grade standards introduced.
Take · A substantive milestone from 'planning' to 'production-line validation.'
Apple官方视频Confidence High
First foldable iPhone Duo (largest screen in Apple history), Watch Series 12 (most accurate heart-rate sensing), and AirPods 5 (open-fit active noise cancellation) launch together.
Take · Device-form-factor battle gains another variable as Apple officially enters the foldable screen race.
Original ↗'AI Procurement Agent' becomes an independent paid category.
淘天 / 京东工业多源Confidence Mid
Taotian launched 'AI Procurement Bao' at CIFTIS (multi-agent handles demand aggregation/sourcing price comparison/order tracking); JD Industrial released a matrix of AI CAD / Design Master / AI Smart Procurement Manager, targeting 63 million small and medium-sized manufacturing enterprises, with over 70 industrial agents deployed in H1.
Take · The category has just been independently priced, and budgets are not yet fixed; the window is about 6-12 months.
03INDUSTRY8 stories
China-U.S. 'distillation' rules dispute turns geopolitical.
美国 NSA/CISA/FBI · 中国商务部官方双源Confidence High
Three U.S. security agencies jointly announced on 9/8, accusing Chinese AI companies of 'industrial-scale distillation' and naming six—DeepSeek/Moonshot AI/Alibaba/MiniMax/StepFun/Zhipu; the Ministry of Commerce responded on 9/9, saying it 'firmly opposes [this] as groundless in fact and law'.
Take · Distillation is securitized by the U.S. for the first time, potentially becoming a new lever for sanctions; track whether it escalates to Entity List inclusion or tighter export controls.
NVIDIA officially announces 15%+ server price hike for 2027
英伟达彭博多源Confidence High
Microsoft, Google and other customers have been notified: Vera Rubin / Grace Blackwell to see price increases starting early 2027; NVL72 rack about $7.8 million, memory's share of BOM rises from 9% to 26%.
Take · Storage price hikes are formally passed through to full-system pricing; compute costs are structurally rising.
Crusoe completes $3 billion financing at a $30 billion post-money valuation.
Crusoe彭博Confidence High
Valuation jumps from $10 billion to $30 billion in 11 months ('10 months' is a misreport); Atreides / Valor co-led, Mubadala participated; also signed an approximately $13 billion 5-year GPU cloud contract with Jane Street.
Take · Primary market goes heavy on compute infrastructure; long-term contracts lock in the demand side.
DeepSeek STAR Market IPO: CITIC Securities has begun due diligence.
DeepSeek21 财经Confidence MidProgress · Day 2
CITIC Securities has begun due diligence but has not yet signed a formal pre-IPO tutoring agreement—upgrading from single-source preparation to hands-on execution by a named securities firm.
Take · The signing of the pre-IPO tutoring agreement is the next hard validation point.
Youdi Robotics' Hong Kong 18C listing debut +153.84%
优地机器人行情数据Confidence High
Closed at HK$36.68, market cap of HK$15.272 billion, raised HK$650 million, with SenseTime among cornerstone investors; Jiazhi and Standard are queuing.
Take · Hong Kong 18C has become the main exit channel for robotics assets, forming a stark contrast with A-share Moore Threads' limit-down on lockup expiry.
PPI today, CPI tomorrow night: the last pricing window before FOMC
市场CME / 市场数据Confidence High
9/10 U.S. PPI, 9/11 CPI (headline YoY expected around 3.4% / core 2.4%); FedWatch puts the probability of a 25bp September rate hike at about 60%; Brent crude breaks $101, 10Y U.S. Treasury yields hit their highest since November 2023, the broad market fell for a third straight day while AI hardware and application-layer funds bucked the trend.
Take · The last pricing window before FOMC, with macro pressure and AI structural trends coexisting.
OpenAI voices support for California AI safety bill
OpenAI官方博客Confidence Mid
Officially supports SB 1119 (minor protection) and other California bills, calls for federal mandatory regulation; SB 813 independent safety assessment passed the legislature on 8/30; 'full list of four bills' not fully independently verified.
Take · Leading vendors shift from resistance to active participation in rule design; California's 30 AI bills face a 9/30 signing window.
Harvey reported to raise $550 million at a $15.5 billion valuation
HarveyTechCrunchConfidence MidSingle-source pending verification
Large single funding round in legal AI sector; single-source, pending verification.
04RESEARCH4 stories
AI autonomous research extends to physical experiments: GPT-5.6 Sol drives quantum chip testing end-to-end
OpenAI × MIT官方案例Confidence MidSingle-party case
GPT-5.6 Sol, driven via Codex, completed the full pipeline of parameter inference / pulse emission / noise-reduction iteration on an unknown 6-qubit superconducting quantum chip; humans intervened only 4 times in 40 measurements.
Take · Single-case example, no peer review; weak-signal scenarios still need human involvement, but the 'automated AI researcher' roadmap is being delivered ahead of schedule.
AA v4.3 leaderboard change: private evaluation weight rises to 45%
Artificial Analysis官方Confidence High
Terminal-Bench rises to 4.0; AutomationBench-AA (a private 657-question set with Zapier) replaces τ³-Banking; after the revamp, Fable 5.1 and GPT-6 Astra tie for first (53), while open-source GLM-5.3 and Kimi K3 tie for first (44).
Take · Not directly comparable with the v4.2 methodology (Fable 56.8 / Astra 54.7)—scores must be cited with the version number.
OpenRouter last week: 115 trillion Token, China 56.72 trillion, surpassing the US for 19 consecutive weeks.
OpenRouter平台数据Confidence Mid
8/31-9/6: 115 trillion Token across all platforms (slightly up from the reported 113 trillion), China share 56.72 trillion.
Take · Call-volume structure is a hard indicator of who is actually being used.
Leaderboard update: Meta Muse Spark 1.3 enters Code Arena WebDev at No. 8.
Arena / LLM Stats榜单Confidence MidSingle-source
Muse Spark 1.3 (Max) priced at $3.50/1M, 30-70% cheaper than peers in the same segment; ChatGPT Images 2.5 (Flare/Sunburst) tops all three Arena image leaderboards; LLM Stats overall ranking rises to No. 4.
Take · Cost-effective models continue to gain on specialized leaderboards.
05INSIGHTS6 stories
Today three threads converge: AI is shifting from a model race to a physical-world and heavy-asset race
前瞻深度版研判Confidence Mid
JD.com three-in-a-row / XPeng production line / NVIDIA price hike corroborate one another; the theme shifts from model-capability narrative to three hard metrics: cost control + compliance + physical deployment.
Take · Outlook for the next 3 months: capital falsification period overlaps with regulatory implementation period.
Structural rise in compute costs; inference-side price cuts + open-source self-hosting as hedge
前瞻深度版研判Confidence Mid
NVIDIA 15% price hike vs DeepSeek cache -60% / Bailing 5.5B active / Cohere megakernel; the 'expensive spring, dig a well' combo holds.
Clear capital overheating signals; entering falsification phase
前瞻深度版研判Confidence Mid
Crusoe 3x in 11 months / Youdi doubled vs OpenAI first disclosure of tens of millions USD cost data for a single project—market starts calculating ROI; Q3 earnings season is the validation point.
Use low-cost multimodal to cut inference costs
技巧深度版建议Confidence Mid
Use Bailing 5.5B activated VL (single-card runnable) to get an invoice / work-order recognition demo working, replacing a 72B solution and cutting costs to 1/10; combine with DeepSeek V4.1 Flash cache-hit pricing to charge per call across the full 'recognition + generation' chain.
Take · The open-source commercialization window is about 3-6 months.
Before compute price hikes, run an 'inference bill checkup'.
技巧深度版建议Confidence Mid
For the recognition portion that can switch to domestic compute / open-source self-hosting, offer compute price comparison and migration/hosting, taking a cut of the savings.
Take · The price hike has been officially announced but not implemented; stack-switching decisions are starting; window 9-15 months.
Two HN hot topics: OECD warns students using AI have worse grades · Anthropic researcher discusses AI risk probability.
社区Hacker NewsConfidence Mid
Today's two hottest views in HN's AI section: research on the negative correlation between AI use and test scores in education is widely discussed; an Anthropic researcher says the probability AI 'could kill everyone' exceeds 10%.
Take · Community sentiment: beyond capability narratives, education and risk topics are heating up in tandem.
Ecosystem Pulseself-updating
News255Skills72OpenHub44Reddit252026-09-02 → 2026-09-10 Star growth：obra/superpowers +90,630★ · mattpocock/skills +82,178★ · multica-ai/andrej-karpathy-skills +67,613★ · anthropics/skills +56,067★ · Shubhamsaboo/awesome-llm-apps +43,732★
✦Action Items3
1. Ride the 'AI Procurement Agent' category waveThis week, review onboarding/API entry points for Taotian's 'AI Procurement Bao' and JD Industrial AI procurement, turn product catalogs into machine-readable formats to grab early traffic slots; next, build vertical middleware for 'procurement list → supplier matching' and take government/enterprise application orders to do delivery and managed operations.
window 6-12 months 2. Use low-cost multimodal to cut inference costsThis week, use Bailing 5.5B activated VL to get invoice/work-order recognition Demo running, replacing the 72B solution; next, layer in cache-hit pricing for per-use billing, then use megakernel-style inference optimization to cut concurrency costs.
window 3-6 months 3. The 'cost lock-in' business before compute price hikesThis week, help teams do a one-time inference bill health check, identify parts that can switch to domestic compute / open-source self-hosting; next, do compute price comparison and migration managed services, taking a cut of savings.
window 9-15 months