01MODEL RELEASES3 stories
Google releases Gemini 3.8 Live as two voice models: ET version scores 82.6 to top AA speech-to-speech leaderboard
Google DeepMind官方博客 + AAConfidence High
9/15 release: Gemini 3.8 Live and Live Extended Thinking, two end-to-end duplex voice models: ET version scores 82.6 to top the AA speech-to-speech quality overall leaderboard (beating GPT-Live-1's 81.5), Big Bench Audio 97.7%, τ-Voice agent tasks 68.6%; core breakthrough is parallel background tool execution during voice conversations and mid-conversation switching across 97 languages.
Take · Pricing is the killer: standard version measured by AA at $0.84/hour (ET version $3.50/hour). With 9/16 StepFun StepAudio 3 (Realtime 98.9%, No. 1 on conversational dynamics leaderboard), it forms a 'two kings in one week'; multimodal voice becomes the next decisive battleground for rankings.
Qwen3.8-27B tops Hugging Face all-time Most-likes
阿里千问HF 榜单 + AIBase/网易/钱江晚报多源Confidence High
9/16 HF Most-Likes leaderboard update: Qwen3.8-27B (Apache 2.0, 27B dense) surpasses FLUX.1, DeepSeek-R1, Kimi-K3, Llama-3 to become the platform's all-time No. 1 by likes—the first Chinese open-source model to top it on community reputation (open-sourced 8/14, about one month).
Take · Perplexity and others have selected it as a local Agent model; overseas community calls it 'local Opus 4.6'; Alibaba's Qwen series has cumulatively open-sourced 460+ models, with over 3 billion global downloads and 300,000+ derivative models. The criterion for open-source competition is shifting from benchmark scores to distribution and reputation.
TypeSafe AI, founded by a ChatGPT co-inventor, launches "decision model" Jev, with a $40 million seed round
TypeSafe AIDCVC 官网 + Business WireVendor self-reported
Diogo Almeida-founded TypeSafe AI launches Jev: it does not generate text, only outputs structured judgments with confidence, and raises a $40 million seed round (led by DCVC).
Take · The "193x faster / 444x cheaper" claims are company self-reported, with no independent evaluation; if true, many enterprise tasks that "only need judgments, not text" would no longer have to pay generation costs.
02PRODUCT4 stories
UBTECH U1 first batch delivery begins: official first disclosure of a 'scissors gap' between orders for 13361 units and 1500-2000 units deliverable within the year
优必选财报电话会 + 南方+ / 阿尔法工场Confidence MidProgress · Day 4
9/16 Shenzhen U1 consumer-grade first-batch deliveries begin (Lite 11.98 万 / Pro 16.98 万 / Ultra 88-99 万元), but the earnings call disclosed actual deliverable volume for the year is only 1500-2000 台; management for the first time anchored 2026Q4 single-quarter adjusted EBITDA as 'having a chance to turn positive', with focus shifting to industrial-side Walker S.
Take · Valuation yardstick shifts from 'counting orders' to 'counting deliveries'. Ledger correction: the 9/14 entry H1 shipments 1.91 万台; under the interim report basis, should be 16123 台.
百度智能云官方 B站Confidence High
Official Bilibili account posted 8 items in one day: industrial agent operating system (collaborative, closed-loop, accountable for results); Agent Infra upgrade for 3 billion industrial agents; Dazi Open Platform (a mature Harness that lets enterprises quickly customize agents, “never takes commissions from partners”); and technical specifications for six core modules of the full-stack domestic Tianchi supernode; Shen Dou said chips, cloud, models, and agents—four layers—will form a mutually amplifying industrial grand cycle.
Take · “Closed loop + accountability for results” aligns with overseas legislation’s “verifiable identity + tamper-proof logs,” as accountability standards for agent platforms converge globally.
Original ↗Salesforce官方 YouTubeConfidence High
The official channel posted several items that day: the main theme 'Hey AI, Meet the #1 CRM'; Enterprise AI Harness for the agentic enterprise; SMB and media breakout sessions; Agentforce from quick start to real ROI, among others.
Take · Along with Baidu AI Cloud, it also touted 'Harness' the same day: orchestration and control layers are replacing the models themselves as the focus of platform competition.
Original ↗康惠股份 RMB 1.72 billion compute long-term order + RMB 1.141 billion server purchase/sale: heavy-asset-for-long-contract model moves downmarket to small and mid-cap listed companies
康惠股份上市公司公告（中证报 9/17 刊载）Confidence High
Wholly owned subsidiary signs a five-year compute service contract with Customer A worth about RMB 1.72 billion (tax included), while signing an RMB 1.141 billion server purchase/sale contract with Company G (a controlled subsidiary of an A+H-listed company) to lock in supply; delivery will be rolled out in batches from the end of 2026Q3, debt-to-asset ratio rises from 69% to about 78%, and new revenue in 2026 is only about RMB 30 million.
Take · Compute outsourcing long-term orders are spreading to small and mid-cap listed companies, but leverage is rising sharply and collections are front-loaded—cash-flow risks at similar targets warrant caution.
03INDUSTRY6 stories
FOMC hikes 25bp to 3.75%-4.00% in unanimous 12:0 vote: first since July 2023, ending the rate-cut cycle
美联储央视新闻 / 每经 / 上证报 / 21 世纪经济报道（多源一致）Confidence High
Dot plot median raised to 4.1% from 3.8% in June; 16 of 18 officials expect at least one more hike this year; SEP revises core PCE up to 3.4% and delays the 2% target to 2029; Warsh's press conference was hawkish, and CME shows about 53% probability of another hike in October. After the decision, Dow -1.21%, S&P -0.45%, and 10Y U.S. Treasury yield rose to 5.008%.
Take · The tightening uncertainty has turned into a fait accompli, giving risk assets an 'expectations-realized' breather; however, 'higher for longer' is now the baseline scenario, and the AI sector is an oversold rebound rather than a trend reversal, with sustainability to be tested by October CPI and FOMC.
OpenRouter weekly leaderboard changes hands: GPT-5.6 Luna tops, US models return to Token top spot for first time in about 20 weeks
OpenRouter官方榜（腾讯新闻 9/16、和讯 9/15 转引）Confidence HighProgress · Day 2 · Metrics must specify the window
As of the 9/13 weekly report window: GPT-5.6 Luna tops with 18.20 trillion Token for the week (+41.1% week-over-week), first for US models since late April; Tencent Hy4 preview is second (16.80 trillion, +14.3%); DeepSeek's four models combined total 21.76 trillion, still No. 1 among vendors; total 127.0 trillion, +10.4% week-over-week.
Take · Real-time leaderboard (as of 9/15) shows Luna 17.4T / Hy4 13.6T—numbers differ by window; citations must specify 'as of the 9/13 weekly report'; if US models hold the top spot for the next two weeks, the 'China's 20-week lead over the US' narrative must be formally rewritten.
Three-track convergence in AI agent governance legislation: US Stop Rogue AI Act + Hong Kong Policy Address + India payment Agent registry
美众院 / 香港 / 印度 NPCIlawler.house.gov / policyaddress.gov.hk / Reuters
Stop Rogue AI Act introduced by bipartisan lawmakers Lawler / Gottheimer on 9/9 and released on the official website on 9/15: directs NIST to develop agent standards within one year—continuous machine-readable agent inventories, verifiable identity and provenance, tamper-proof operation logs, real-time behavior monitoring and permission revocation, mandatory for federal contractors; Hong Kong's 9/16 Policy Address proposes seven areas of action for AI risk governance and next year will release society-wide AI Agent safety management guidelines; India's NPCI payment Agent unified registry as background.
Take · Regulatory focus shifts from 'model capability ceilings' to 'agent behavior control'; lawmaker statement explicitly links it to the July incident where 700 OpenAI agents breached Hugging Face—a rare common denominator across China, the U.S., and Europe.
ByteDance's AI pharma unit spins off as Anew Labs: $290 million first round, $1.5 billion post-money valuation
字节跳动Reuters 9/16 + Moneycontrol / TNW / DealroomConfidence Mid
The spin-off closed a $290 million first round the same day, led by HSG (Sequoia China) / IDG / Hillhouse, with FiveYuan co-leading, and Gaorong / Primavera / Boyu / Shanghai Future Industry Fund participating; ByteDance retains a 56% stake.
Take · 'Big Tech spin-off + independent financing' becomes a new channel for monetizing AI businesses, directly raising capital levels in AI × high-value science.
Capital snapshot: Euclyd $231 million Series A (non-GPU inference architecture), CADDi $114 million Series D
Euclyd / CADDiCNBC 独家 + Fortune / WSJ
Netherlands-based Euclyd closed a $231 million Series A, co-led by Samsung and Somerset / EQT Scaleup / Innovation Industries; former ASML CEO Peter Wennink serves as chairman; focuses on non-GPU inference architecture and plans physical chips in 2028; manufacturing drawing data platform CADDi closed a $114 million Series D at a $1.2 billion valuation (Toyota Woven Capital / Salesforce Ventures participated; Subaru / Kawasaki / Yanmar are customers).
Take · '100x energy efficiency' is the company's forecast basis; the stronger general models become, the more valuable proprietary industry operational data is.
Two pending-verification governance developments: the draft 'Common Principles' charter from the UK Royal AI Summit, and the launch of the first periodic inspections of high-risk systems under the EU AI Act
英国王室 / 欧盟聚合简报源Confidence MidSingle-source pending verification
The UK Royal AI Summit is being held this week at Dumfries Estate, as a Ditchley 'Common Principles' charter draft surfaces (Dario's slowdown essay partly counted toward ledger progress); the EU AI Act's first periodic inspections of high-risk systems begin (24 member states, HR screening / credit scoring / triage), with aggregated briefing sources only and no official announcement seen.
Take · Neither has official text, so both are on hold under the 'report after official confirmation' discipline; if they land, they will form a four-track resonance with the US / Hong Kong / India, and agent behavior control will become the global default regulatory model.
04RESEARCH4 stories
Mozilla 'State of Open Source AI' v1.1: open-source vs closed-source gap narrows to 4.4 months, revenue scissors gap 4% vs 96%
Mozilla官方报告（Ars Technica 独家预发 / IT之家）Confidence High
Based on METR time-horizon fitting: China's best open-weight model is only 4.4 months behind the US frontier closed-source model (Kimi K3's intelligence index is only 3 points lower than Claude Fable 5, with cost about 30%; GLM 5.2 is within 1 point of Opus 4.7/4.8, with per-task cost about 5x lower); the closed-source premium exists only in the 8-12 hour task range.
Take · In OpenRouter's August Token volume Top10, 8 are open-weight (7 Chinese-made), but open models account for only 4% of model-layer revenue (closed-source 96%; Linux Foundation paper, May-September 2025 window) — 'affordable ≠ profitable'; the revenue lag itself is the next variable to watch.
Tsinghua 'post-95' Chen Yongchao's Chaoyan Intelligence raises nearly RMB 400 million in 3 months, focusing on RSI recursive self-improvement
超衍智能央广网 / 证券时报 / 科创板日报 / 贝壳财经多源Confidence High
Company founded in June 2026; on 9/16 announced cumulative funding of nearly RMB 400 million across its angel round + angel+ round: angel round led by IDG + Xinglian Capital + XtalPi, angel+ round co-led by Shenzhen Capital Group / Zhongguancun Science City / Shanghai Future Industry Fund; uses 'research trajectories' as training signals, with research automation as first landing point, and claims to build a 'network of 100,000 AI scientists'.
Take · The catch-up logic shifts from 'chasing the frontier' to 'improving speed itself'; deep participation by state-backed patient capital in early rounds also means the capability-jump risk of RSI (black swan B2) is starting to be priced in.
Benchmark methodology correction: AA voice leaderboard 82.6 is for the ET version, comparison baseline 81.5; OpenRouter weekly report and real-time leaderboard figures cannot be mixed
Benchmark 口径AA / OpenRouter 修正Confidence HighMethodology discipline
AA notes the voice-to-voice quality leaderboard is a single trial and noisy; the 82.6 score cited in this article is for the Gemini 3.8 Live Extended Thinking version, while the standard version has a separate pricing basis; OpenRouter has two sets of figures for the same period: weekly report (Luna 18.20T) and real-time leaderboard (Luna 17.4T).
Take · This week, two widely circulated figures both show 'version/window mixing'; cited scores and shares must include version numbers and cutoff windows, otherwise conclusions can reverse.
Canada Float report: enterprise AI spending up 12x in two years, 47% of enterprises purchasing AI, average 2.71 vendors per enterprise
FloatFloat 报告（1 万+ 企业真实流水）Confidence Mid
Based on 10,000+ enterprises' real transaction data (not a census): enterprise AI spending grew 12x in two years, 47% of enterprises are already purchasing AI products, each using an average of 2.71 vendors.
Take · 'Multi-vendor coexistence' rather than single-platform lock-in—mutually corroborated by Mozilla's revenue scissors gap: on the enterprise side, they can afford it and use it heavily, but revenue has not yet concentrated among a few vendors.
05INSIGHTS5 stories
Today's main line: rate hike lands × scoreboard generational shift, AI enters an 'operating discipline period'
前瞻深度版研判Confidence Mid
FOMC hikes for the first time in more than three years (12:0 unanimous, +25bp), tightening the macro reins; same day OpenRouter weekly rankings change hands, voice leaderboard has dual kings from China and the US, agent legislation resonates on three fronts—the scoreboard is shifting from 'model benchmark scores' to 'Token consumption × behavioral compliance'.
Take · Industry shifts from an arms race period to an operational discipline period: capability narratives give way to three hard metrics—cost, compliance, and delivery; the next 3 months are an overlap of a capital falsification period and a regulatory implementation period.
Valuation yardstick shifts: from 'counting orders' to 'counting deliveries'
前瞻深度版研判Confidence Mid
优必选 orders of 13361 units vs 1500-2000 units deliverable within the year were officially quantified for the first time; long-term compute orders represented by 康惠股份, by contrast, use leverage to secure long-term contracts (debt-to-asset ratio 69%→78%, about 3000 万 in new revenue for the year).
Take · 'Contract amount' no longer equals 'recognized revenue': robots look at monthly deliveries, compute looks at payment-collection pace; both types of targets have entered a fulfillment-verification period.
The open-source monetization gap is the most certain opportunity area right now
前瞻深度版研判Confidence Mid
Open weights account for about 8 成 of Token volume but only 4% of model-layer revenue (closed-source 96%); Qwen3.8-27B tops HF Most-likes by community reputation, confirming distribution capability has become the first asset of open-source models.
Take · Gap = servitization space: private deployment + industry fine-tuning + cost routing—the direct business of 'affordable but someone must deliver'.
Capture the $0.84/小时 voice foundation dividend and build vertical Agent
技巧深度版建议Confidence Mid
Foundation prices were halved within one month, making phone order-taking / appointment front-desk scenarios that did not pencil out a year ago now viable; use multi-model routing to turn the price war into your own cost moat.
Take · Window: about 12-18 months (before big tech enters vertical SaaS directly); first use a 'restaurant reservation phone Agent' to build a demo in one week, then replicate at 'X yuan per order taken'.
Two official points: Claude Cowork and chat merged into the same Claude; quoting Mustafa Suleyman
社区Simon Willison 官方博客Confidence Mid
Simon Willison 9/16 two updates: Anthropic merged Claude Cowork and chat into the same Claude; also reposted a passage from Mustafa Suleyman (Quoting Mustafa Suleyman).
Take · The product entry point consolidation happens in sync with the enterprise-level Agent narrative; Suleyman's quoted original text can be found at https://simonwillison.net/2026/Sep/16/mustafa-suleyman/, and can be kept on file as first-hand market-side opinion.
Ecosystem Pulseself-updating
News509Skills72OpenHub44Reddit302026-09-02 → 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. Agent compliance audit: build the 'Agent three-piece set' during the standard finalization periodFor AI companies going overseas and teams taking government/enterprise contracts, provide a three-piece SDK / middleware set and compliance self-check service covering 'Agent identity + tamper-proof operation logs + behavior auditing.' Path: break down the three documents NIST / Hong Kong / NPCI into a Chinese comparison checklist (free report) → deliver lightweight audit middleware to 2-3 lighthouse customers → refine into a standard product, bound to the 'federal contractor mandate' overseas expansion chain.
window 6-12 months (NIST standard-setting period = alignment period; once the standard is released, it becomes a red ocean) 2. Voice-native vertical Agent: capturing the $0.84/hour base-layer dividendDon’t build the base layer—build delivery: AI phone order-taking / appointment front desk for SMBs (restaurants, home services, clinics), AI voice practice partner, charged on outcomes. Path: use a low-cost domestic base layer to build a Restaurant Reservation Phone Agent demo within one week → validate connection and conversion rates with 3 pilot merchants → replicate to 50 merchants on a per-order X yuan basis → add parallel tool calling to connect directly to order systems.
window 12-18 months (before big tech enters vertical SaaS themselves) 3. Open-source monetization gap: private deployment + Token cost routingCapitalize on the gap where open source accounts for 80% of Tokens but only 4% of revenue—provide data-sensitive SMEs with Qwen3.8-27B-level private deployment + industry fine-tuning; provide high-consumption teams with open-source hybrid routing, cutting inference bills by 30-50%. Path: publish an Open Source vs Closed Source Selection Guide + Cost Calculation Sheet → take 5 private deployment orders (27B can run on a single card) → turn legal / e-commerce / education industry templates into subscriptions + routing commissions.
window About 12 months (pitch is strongest before the gap narrows to within 3 months)