01MODEL RELEASES4 stories
Xiaomi MiMo-V2.6-Pro (1T total params / 42B active) tops the open-weight leaderboard: RL training under 6 days, about 750,000 trajectories, cost about $2.62 million
小米Latent Space AINews 2026-09-22 + 小米 MiMo 官方文档（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Xiaomi MiMo-V2.6-Pro is a trillion-parameter, 42B-active omni-modal reasoning model; official documentation states: RL training under 6 days, 30 steps each, cumulative about 750,000 trajectories; Flash version cost about $850,000, Pro version about $2.62 million (total about $3.5 million); overseas engineering community Latent Space (AINews) calls it the new strongest open-weight model, while another source gives Artificial Analysis Intelligence Index 46.32 and per-task cost $0.13.
Take · Pushing the acquisition cost of the 'open-source frontier' down to the million-dollar level shifts pricing power over capability gaps from capital to data and post-training design; together with StepFun Step 5 on the same day, it forms a 'Chinese open-source double header,' directly rewriting the in-house vs external procurement decision line—what should be recorded in reports is '6 days + $2.62 million,' not total parameter count.
StepFun releases Step 5 Preview: 600B total params / 27B active sparse MoE, 1 million Token context; official says October open-source
阶跃星辰量子位 2026-09-22（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Sparse MoE architecture, 600B total params with only 27B active (activation ratio about 4.5%), supports 1 million Token context; official says it is on the 'Pareto frontier' of the AA leaderboard for Intelligence Index and per-task cost, ranking in the open-source top three, with open-source timing stated as October.
Take · Sparsification separates the 'capability ceiling' from the 'bill per call,' so comparing models should shift from looking at total params to looking at capability/cost ratio—model selection discipline becomes drawing your own cost curve first, then choosing models on that curve; and the October open-source is a publicly verifiable commitment.
Grok 4.7 released, input priced at $2 / million Token
xAI（Grok）界面新闻 2026-09-21（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidSingle-source pending verification
Local intelligence log records the day's pricing anchor: Grok 4.7 released with an input price of $2/million Token (original media: Jiemian News, 2026-09-21, single source).
Take · Read this alongside Step 5's 'Pareto frontier' and open-source low pricing, and it is a three-way pincer on the floor of the same price range: the yuan/million Token price anchor is converging toward the $2 order of magnitude, and any cost model based on old unit prices needs to recalculate its denominator.
HONOR and Alibaba jointly build five vertical-domain models: device makers and cloud providers co-build vertical-domain agents
荣耀 / 阿里云环球网 2026-09-22（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidSingle-source pending verification
Local intelligence log: HONOR and Alibaba jointly build five vertical-domain models (Huanqiu.com 2026-09-22, single source), aligning with Qualcomm's same-day 'agent-centric' on-device narrative.
Take · On-device + vertical domain is currently the most realistic path to deployment for 'agents': who provides the vertical-domain models and who controls the device entry point directly determines the revenue-sharing structure on the device side—this can change procurement relationships more than launching another general-purpose model.
02PRODUCT4 stories
Qualcomm's sixth-generation Snapdragon 8 Super Ultimate Edition: focusing on "agent-centric," with on-device support for 30-billion-parameter models and prefill throughput above 330 Token/s
高通IT之家 2026-09-23（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Officially claimed to be the world's fastest mobile CPU, it can run 30-billion-parameter-class models on-device with prefill throughput above 330 Token/s; Xiaomi, vivo, OPPO, RedMagic, and REDMI will be among the first to support it; CEO Amon says the industry is shifting from "phone-centric models" to "agent-centric models."
Take · For the first time, "agent" is written as a benchmarkable on-device hardware metric, but what is given is prefill, not decode—what truly determines perceived experience is decode throughput, memory bandwidth, and throttling and temperature after ten minutes of continuous inference. The acceptance checklist should add exactly these three items, not TOPS.
Runway📹视频库《AI官方视频监控台账》2026-09-23（YouTube·Runway）Confidence MidOfficial
Runway's official channel releases DIFFUSE; the ledger entry only provides the title and link, the video body was not obtained, and specific capabilities and parameters are subject to the official release.
Take · A new product on the video-generation side, directly corresponding to our video factory's visual supply, and can immediately be added to the "asset generation engine" comparison checklist.
Original ↗GitHub📹视频库《AI官方视频监控台账》2026-09-23（YouTube·GitHub）Confidence MidOfficial
GitHub's official channel released a Copilot Day-themed video, setting the tone for agentic engineering; the ledger entry only provides the title and link, and the body was not retrieved.
Take · The code-hosting side's official positioning on “agents writing code” is a phase marker of coding agents moving from assistance to hosting—questions like how review processes change and who is responsible for merges will surface first on this front.
Original ↗TypeSafe AI / Simon Willison📹视频库《AI官方视频监控台账》2026-09-23（官方·Simon Willison，已回读原文）Confidence HighOfficial
He re-read the original to verify: after llm install llm-typesafe and llm keys set typesafe, running llm -m jev -s 'decision question' can handle three task types—yes/no questions (the confidence of the example output {"type": "noul", "noul": 0.99}), choice multi-classification (e.g., routing tickets to billing/technical/other), and score scoring (e.g., assessing a report's reproduction completeness on a three-point scale).
Take · Replace the most expensive step (human judgment) with a machine-checkable contract: probabilities, enums, scoring tiers. Our judging panel, figure review gate, and plagiarism initial screening can all be adapted directly to this interface, at far lower cost than training another model.
Original ↗03INDUSTRY6 stories
Alibaba Cloud Apsara Conference: T-Head Zhenwu V900 self-developed AI chip released, 5 trillion to 10 trillion parameter large model previewed, 2032 data center target over 20GW
阿里巴巴 / 平头哥IT之家、通信世界（网易）2026-09-22（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Zhenwu V900 combines with self-developed ICN Switch to form a supernode, with native memory semantics and unified memory addressing, and thousand-card full-bandwidth high-speed interconnect (another account says a single cluster can scale to 500,000 cards); officials also previewed a 5 trillion to 10 trillion parameter-class large model and set Alibaba Cloud's 2032 global data center scale target at over 20GW, shifting the framing from 'how big is the model' to scaled value creation in real-world scenarios.
Take · This is the only release this week that reported all four layers—chip, interconnect, model, power—at once. The signal is readable: the competitive unit for compute self-reliance is shifting from single-card performance to interconnect bandwidth + power quota + cluster scalability; pricing and scheduling should switch to this denominator; and 'trillion-parameter' is still only a preview—it will not count as delivered until node and inference costs arrive.
Compute infrastructure retrenches on two fronts: OpenAI and Anthropic shift to 20–30MW small data centers; about 45 US projects worth $68 billion have been halted or delayed due to community opposition
OpenAI / Anthropic / 美国数据中心项目Tom's Hardware 2026-09-22、界面新闻（Data Center Watch）2026-09-21（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Because construction of gigawatt-scale hyperscale data centers is behind schedule, the two companies are instead seeking 20–30MW small facilities to handle current compute demand from millions of GPUs; meanwhile, Data Center Watch counts that from April–June 2026, about 45 data center projects worth $68 billion were halted or delayed, and about 30 state legislatures have proposed or passed regulations involving siting, power, and water use.
Take · The constraint has shifted—the bottleneck has gone from 'can you buy chips' to 'where does power come from, who approves the land, and will residents agree.' Compute certainty is lower than expected, meaning any pricing or SLA premised on 'unlimited compute and continuously falling costs' lacks buffer; under the newsvendor model, starting small-batch plus keeping an expansion option is the minimum retractable cost.
AMD's total market cap tops $1 trillion for the first time, becoming the fourth U.S. chip company to join the trillion-dollar club
AMD财联社 / 36氪 2026-09-22（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Its U.S. share price broke above $613, with intraday gains once at 9.5% (another measure: closed at $615.52, up 9.95%), becoming the fourth U.S. chip company with a market cap above $1 trillion.
Take · This forms a tension with the same-day '45 projects blocked + two leading players retreating to small server rooms': capital flows are accelerating while physical supply is putting up roadblocks. When the capital narrative runs more than two years ahead of cash flow, what to watch is not funding news but actual renewal rates and grid-connection progress.
Kimi K3 joins Amazon Bedrock: a Chinese large model is for the first time distributed to the world's three major clouds under a revenue-sharing model
月之暗面 / 亚马逊 AWS36氪 2026-09-22（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Global enterprise developers can directly call Kimi K3 on Bedrock; Moonshot AI is advancing revenue sharing with multiple overseas cloud vendors; K3 was previously listed on Alibaba Cloud Bailian.
Take · When the settlement model changes, the nature changes—from 'selling API / one-time licensing' to 'appearing on the counterparty's bill and sharing revenue'; the channel party shifts from buyer to agent, and for the first time sales and payment-term incentives align with our revenue. But for revenue sharing to work, first clarify what counts as billable revenue, how trials and refunds are handled, and who holds audit rights; otherwise disputes will grow in reconciliation.
Entry and monetization diverge: Meta agent Muse tops Apple US App Store, driving A-share concept stocks broadly up over 2%; meanwhile ByteDance Doubao general Session team size is expected to halve
Meta / 字节跳动证券时报 2026-09-23、晚点 LatePost / IT之家 2026-09-22（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Muse, launched in early September, quickly overtook ChatGPT and Claude to top the US charts, driving gains of over 2% across A-share AI office, virtual human, and agent concepts (another account: topped the free chart within three weeks of launch, maintained 4.9 stars with 25,000+ reviews); in the same week, Doubao's once-largest general Session team (once about 50 people) began staff cuts, with some transferred to Doubao commercialization and Feishu, and the conversational product post-training team was also downsized, amid a commercialization bottleneck after surpassing 200 million DAU.
Take · Only reading both together gives the full signal: distribution ranking can be operated; retention and per-user monetization are the expected value—No. 1 on the chart is a distribution dividend, not a product victory. Dialogue is the previous generation's interface dividend; a closed loop that can complete tasks and settle payments is this generation's execution dividend; products that only do conversational form should now answer 'at which step does the transaction happen?'
Two lines: rules and trust: Zhipu responds to ZCode data privacy controversy—will soon officially open-source and invites CAICT to audit; OpenAI urges US to lead global standards for frontier AI
智谱 / OpenAI澎湃新闻（36氪转载）2026-09-22、新浪财经 / 36氪 2026-09-22（本机 20 bot《每日AI情报 2026-09-23》转述）Confidence MidMultiple sources agree
Zhipu says open-sourcing is the first step to address data retention concerns, will focus on auditing data retention, and establish a regular product security vulnerability reporting mechanism; same day, OpenAI argued for defining which AI incidents must be reported, clarifying how countries should cooperate, and improving access to compute.
Take · The two things are the same move—replacing 'trust me' with 'verify me.' The value of a credible signal comes from a cost rivals cannot easily imitate at low cost (open-sourcing + third-party audits require real money), so it is informative; caveat: checking only data retention and not the full chain will backfire into 'performative transparency,' as precedents exist.
04RESEARCH2 stories
HF Daily Paper: Prediction-Powered Smoothing and Validation for Disaggregated AI Evaluation — treating the evaluation set as a finite population for stratified smoothing
HF Daily Papers（arXiv 2609.20758）HF Daily Papers 2026-09-23 抓取（论文 2026-09-17，arXiv 2609.20758；本机 20 bot《每日AI情报 2026-09-23》坎流转述）Confidence MidSingle-source pending verification
The paper directly addresses 'full annotation is too expensive and sampling conclusions are unstable': it treats the evaluation set as a finite population and uses prediction enhancement for stratified smoothing; local intelligence therefore recommends fixing the evaluation protocol into a three-part set — sample size n, confidence intervals, and A/B significance thresholds — and labeling all conclusions with n<30 as 'for reference only.'
Take · Not academic interest but cost saving: our leaderboard aggregation, judge-panel scoring, and image-review gates are all essentially sampled decisions. Writing n / confidence intervals / significance thresholds into the process is equivalent to adding a verifiable boundary to every 'looks like it got better' conclusion.
Fireworks AI📹视频库《AI官方视频监控台账》2026-09-23（YouTube·Fireworks AI）Confidence MidOfficial
Fireworks AI's official channel releases a new index; the ledger source only provides title and link, the body was not obtained; index methodology, samples, and weights are subject to the official release.
Take · Evaluation methodology is model-selection pricing power: the emergence of a 'specialized' index beyond general leaderboards shows that evaluation segmented by task domain is becoming a common language. It can serve as an external anchor for our 'choose models by task,' but first confirm whether its methodology is vendor-defined before citing it.
Original ↗05INSIGHTS6 stories
Add a hard metric to the edge AI acceptance checklist: throughput retention rate and surface temperature after 10 minutes of continuous inference
本机 20 bot 情报共识（端侧验收）☯九元素 20 bot《每日AI情报 2026-09-23》（悦泽视角）Confidence MidLocal intelligence consensus
Intelligence quote: Vendors book the cloud costs they save into their own cost account, and book charging count and thermal throttling into users' experience account, so battery-life figures are never seen at launch events; in contrast, on the same day 后摩智能 confirmed that its next-generation edge chips will adopt 3D compute-in-memory, precisely to find a solution for this unreconciled ledger of 'heat and power'.
Take · The reason users turn off edge AI is never 'not smart enough,' but heat, battery drain, and slowdown. This metric predicts retention better than TOPS, and directly determines how our edge pilot should be accepted.
Minimal edge vs cloud A/B: run the same task n=30 times on each, record P50/P95 first-token latency, per-run power consumption, and offline availability rate
本机 20 bot 情报共识（实验口径）☯九元素 20 bot《每日AI情报 2026-09-23》（渊策视角）Confidence MidLocal intelligence consensus
Criterion from the intelligence: if edge P95 first-token latency is <800ms and offline availability rate is >95%, then an 'edge-first' architecture holds; it also notes that 330 Token/s is prefill, not decode, and what truly determines experience is decode throughput and memory bandwidth.
Take · Turn 'whether edge can really work' into an experiment that can yield a conclusion within seven days, not a debate. Only if the criteria are fixed first can conclusions be reproducible—this is exactly the shape our acceptance discipline should have.
Evaluate AI apps only by 30-day retention and active-open share, not download rankings; for product builders, reverse it—secure the default slot first.
本机 20 bot 情报共识（留存指标）☯九元素 20 bot《每日AI情报 2026-09-23》（悦泽、昆仑视角）Confidence MidLocal intelligence consensus
Intelligence judgment: App store default and recommendation slots make decision cost zero; users don't need to like it, only to tap it casually; reading 'No. 1 on the chart' as a demand surge or treating it as a valuation anchor both mistake momentary noise for expected value.
Take · This points the way for both sides at once: when reviewing others' products, look at retention and active opens, not ranking position; when building your own entry point, choice architecture (default placement) is far cheaper than persuasion.
For model selection, budget by the 'open-source price for the same capability in 6 months,' not today's flagship quote; compare models not by total parameters but by capability/cost ratio.
本机 20 bot 情报共识（选型纪律）☯九元素 20 bot《每日AI情报 2026-09-23》（风行、慧师、厚坤视角）Confidence MidLocal intelligence consensus
Intelligence framing: when an equivalent intelligence index can be reproduced at extremely low cost, the premium of closed-source flagships can only be supported by trust and compliance; the capability-leadership portion will depreciate quickly; model selection docs should add two columns: 'distribution channel availability + compliance self-certification cost.'
Take · A day late in changing the budget means paying another round of premium. This ruler turns 'saving money' from a negotiation tactic into a process parameter, and can be written directly into our procurement and self-build evaluation templates.
SLA should specify elastic degradation strategy and rate-limiting priorities; for self-built compute, list “grid-connection time” as the top risk and add regional redundancy.
本机 20 bot 情报共识（算力条款）☯九元素 20 bot《每日AI情报 2026-09-23》（悦泽、巽风、岱衡视角）Confidence MidLocal intelligence consensus
Intelligence alert: compute unavailability will reach users in the most basic ways—rate limits, queues, price hikes; money already invested in giant infrastructure is sunk cost, and continuing to wait for it is looking backward; under real options, make small investments by zone and keep an exit at each stage.
Take · Translating external risk into contract language is cheaper than explaining after the fact. This item relies on no forecasts and should go into the template today.
Simon Willison（引 @therealcornpop）📹视频库《AI官方视频监控台账》2026-09-23（官方·Simon Willison，已回读原文）Confidence HighOfficial
The original quotes a TikTok creator’s critique: the key is not the common AI tone (“not X but Y,” three-part structure, fragmented staccato feel) but the hollowness of the content—lacking your own voice, so listeners can tell you have no point of view on what you’re discussing.
Take · This directly maps to our voiceover and content line: what the audience hears is judgment, not tidy sentence patterns. “Opinions must be provided by humans; voice can be replicated” is exactly why we separate the two.
Original ↗Ecosystem Pulseself-updating
News600Skills66OpenHub44Reddit252026-09-02 → 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 Items4
1. On-device AI acceptance checklist leads with three added columns: “10-minute throughput retention rate + surface temperature + offline availability rate.”On on-device-related acceptance forms, fix three hard metrics: throughput retention rate after 10 minutes of continuous inference (vs. first minute), chassis surface temperature, and offline availability rate; any report giving only TOPS or parameter scale must be marked “incomplete scope.” Advanced: run one minimal A/B—same agent task on-device and cloud, n=30 each, record P50/P95 first-token latency, per-run power consumption, and offline availability rate, with criteria fixed in advance (if on-device P95 first token <800ms and offline availability rate >95%, then “on-device-first” holds).
window Immediate - 1 month 2. Model selection document adds two columns: distribution channel availability + compliance self-certification cost; budget is based on “same-capability open-source price in 6 months.”Selection template adds two fixed new columns—distribution channel availability (whether it can enter cloud marketplaces / whether it can revenue-share rather than only API), compliance self-certification cost (audit, data retention statement, vulnerability reporting mechanism); unit price budget is no longer based on current-quarter flagship quotes but estimated at “same-capability open-source price in 6 months,” and reference anchors are recorded alongside (Grok 4.7 input 2 USD/million Token, Step 5’s Pareto frontier scope, MiMo-V2.6-Pro’s cost disclosure). Advanced: run one round of open-source comparison for own core tasks (Step 5 / MiMo-V2.6-Pro vs current model, fixed task set n=30, capability-cost curve in 7 days).
window Immediate - 2 weeks 3. Compute and service contracts add two items: flexible degradation clauses, and regional redundancy plus revocable timing points.External service SLAs should specify flexible degradation strategy and rate-limiting priorities (trigger conditions for queueing, rate limiting, and priority tiering); for self-built or long-leased compute, list “grid-connection time” as the No. 1 risk item (not electricity unit price), build redundancy by region, and split investment into three phases—validation phase—ramp-up phase—scale phase—each with exit gates and revocable timing points. Basis: same-day 20–30MW pullback + 45 projects blocked + 30-state legislation all point the same way; compute certainty is lower than expected.
window Immediate - 1 month 4. Turn “human judgment” steps into machine-checkable contracts: probability / enumeration / scoring bands.Following the llm-typesafe interface shape, convert human adjudications such as judge-panel scoring, image-review gates, and initial plagiarism checks into structured calls: yes/no returns a probability, multiclass returns an enum, ratings return score bands and criteria, and raw outputs are retained in a ledger; evaluation-type conclusions also lock in sample size n, confidence interval, and significance threshold, with n<30 always labeled "for reference only". Advanced: wire the three decision types into existing gate scripts to form replayable decision logs (replay = recheck).
window Immediate - 1 month