01MODEL RELEASES5 stories
DeepSeek V4.1 Flash: DeepSWE up to 77.0%, per-task cost under $0.2, Search/Fetch API opens for the first time
DeepSeek亿邦动力 + 明见交叉（DeepSWE 口径未公开）Confidence MidNew topic
Official figures put per-task cost under $0.2, while competitors cost $2–10+ per task with no higher accuracy; the two Search/Fetch APIs are open externally for the first time at $0.75 per thousand calls. 明见 cross-checked: median per-task cost about $0.18–0.19, based on 7498 real sessions, with composite net improvement 5.13%.
Take · The savings are not on the inference bill; the retrieval layer is directly commoditized—the marginal cost of self-built RAG is undercut by vendors' low-priced bundles, and the middle layer is thereby pocketed by vendors. Trade-off discipline: 0.18–0.19 is a cross-derived estimate, not an independent second-source measurement, so it is not upgraded to 'High'; DeepSWE samples and methodology are not disclosed, so treat it only as a 'reading worth re-verifying'.
Mistral open-sources Large 4 weights: 1T parameters, focused on Token efficiency for coding agents
MistralIT之家Confidence MidNew topic
Trillion-parameter-level weights are open, positioned around Token efficiency in coding-agent scenarios.
Take · Europe's open-source camp makes its first head-on push at the trillion-parameter level; combined with two OCR open-source releases the same day, this directly raises the opportunity cost of in-house mid-sized models—with 'big and comprehensive + small and fast' both open-sourced at once, the most uncomfortable are the middle batch of in-house models.
Open-source OCR blossoms at both ends on the same day: 文心 Unlimited OCR 2 highlights 'unlimited' page counts, and 小米 0.8B outputs Markdown/JSON directly
百度 / 小米IT之家Confidence HighNew topic
Baidu open-sources ERNIE Unlimited OCR 2 (focusing on "unlimited" page count), and Xiaomi open-sources a 0.8B OCR small model (directly outputs Markdown/JSON, on-device structured output).
Take · OCR is the entry layer for structured data; both companies open-sourcing this layer effectively pushes the price of the first step in enterprise document pipelines to near zero; the on-device 0.8B direct structured output also addresses the compliance selling point of "data does not leave the intranet."
DeepSeek Cyber cybersecurity standalone model enters limited gray release
DeepSeek亿邦动力 / IT之家Confidence MidNew topic
A general model vendor has launched a standalone model for cybersecurity, currently in small-scale gray release.
Take · The first move by a general model vendor into vertical industries; the commercialization ladder remains to be seen: vertical models sell "scenarios where they dare to take responsibility," not scores. This is only a directional clue; there are no public readings on gray-release scope or effect.
QbitAI says new Claude models have been exposed: new Fable and new Opus
Anthropic量子位（转述，无第二源）Pending verification
QbitAI reports that a new generation of Claude models has been exposed, with naming pointing to "new Fable" and "new Opus".
Take · No official documents can verify the naming or release timing; treat as 'pending verification' and do not include in any conclusions. The reason to include it in the model: when a frontier lab changes naming, it often comes with a product-line split (capability tier vs price tier), making it worth an observation point.
02PRODUCT LAUNCHES4 stories
Apple Intelligence in mainland China to integrate Alibaba Qwen: developer Beta testing starts November 4, official rollout in December with iOS 27.3
Apple × 阿里千问IT之家（20 份情报全数收录，多稿互补）Confidence HighNew topic
Siri, Writing Tools, and Visual Intelligence can directly invoke Qwen; real-time translation expands to AirPods 4 noise-cancelling version/Pro 2, with Qwen account login supported. This round does not include Siri AI or WWDC 2026 next-generation features; devices limited to iPhone 16 and later, and iPhone 15 Pro series.
Take · Entry point changes hands: Chinese-language intelligence capabilities on mainland China iPhones are handed to a domestic model; this is a 'compliance minimum viable version'—a pilot with limited authorization that can be withdrawn at any time. The clearer the boundaries, the more it shows this is a phased handover rather than a one-time switch; for teams building on-device applications, Siri's callable boundary is the entry point for a new round of distribution positions.
Tencent WorkBuddy has accumulated 9 billion task requests: enterprise edition deployed in over 50 industries; Buddy One and Hunyuan Hy4/Hy4 flash launched at the same event
腾讯腾讯新闻／央广网 + 财经头条（厂商口径）Confidence HighNew topic
Cumulative 9 billion task requests completed; enterprise edition covers 50% of China Enterprise 500 and 60% of China Top 100, with over 2,000 ecosystem partners (over 60% first-time cooperation); at the same event, Buddy One and Hunyuan Hy4/Hy4 flash were released (Hy4 introduces visual understanding for the first time).
Take · Agent gives an auditable volume for the first time — but only the first of the three dashboard columns. Based on roughly 245 days from the March launch to end-October, average daily task requests are about 36.73 million (Kanliu estimate); the whitepaper did not disclose paid conversion rate or ARR per customer, and 'task requests' ≠ 'paid tasks' — mixing these definitions is the easiest way to fool oneself.
Zest's new AI agent: uses SMS to find restaurants and complete reservations
ZestTechCrunch（官方标题为准）Confidence HighNew topic
TechCrunch reports that Zest launched a new AI agent that can find restaurants and make reservations on users' behalf via SMS.
Take · Another TechCrunch item the same day offers a sobering read: DoorDash says it received only 'a few hundred' orders from AI agents. The two must be read together — capabilities are spreading, but settlement volume is still very early; any claim that 'Agents have taken over the consumer side' lacks data support.
Google Cloud官方视频（YouTube·Google Cloud）Confidence MidNew topic
Google Cloud's official channel released CodeMender, positioned to directly fix vulnerabilities at the GitHub Pull Request stage.
Take · Security tools are shifting from 'scan and alert' to 'submit patches,' meaning they must enter CI to count as delivered. This also changes the evaluation standard: not how many vulnerabilities were reported, but how many patches were merged and how much rework was required.
Original ↗03INDUSTRY6 stories
Kling first round closes at RMB 19.802 billion: down RMB 645 million from the RMB 20.447 billion cap, at a post-money valuation of about USD 18 billion
可灵 AI（快手）科创板日报（新浪转载）Confidence HighNew topic
Actual proceeds were RMB 645 million below the cap (3.15% of the cap); the shortfall came from two BlueFive entities withdrawing and Kingdee and MetaX taking over nearly RMB 800 million, while Kuaishou remains controlling shareholder.
Take · Primary-market disagreement over video generation has been written into the subscription book for the first time. Kanliu calculation: RMB 19.802 billion ÷ about RMB 122.7 billion valuation (USD 18 billion × 6.82 exchange rate) ≈ 16.1% dilution—received amount is actual; dilution rate is inferred (the 6.82 exchange rate is an assumption). More worth tracking than the valuation itself is the act of someone withdrawing.
Three linked compute credit-chain items: Anthropic expected to get a junk rating, NVIDIA in talks to guarantee Nscale, OpenAI reported to plan to buy Google TPUs
算力信用链（Anthropic / 英伟达 / OpenAI）财联社／智通财经／智东西引 The Information（三条均单源·知情人士）Confidence MidNew topic
Blackstone has told investors in the USD 63 billion chip financing: Anthropic is expected to receive only the higher tier of junk ratings from major rating agencies after listing; NVIDIA is in talks to provide major credit support for Nscale's large AI data center; OpenAI is reported to plan to buy Google TPUs.
Take · The chip seller is providing a credit backstop for the buyer, while the buyer is seeking a second source for itself—two spread scenarios for the same principal are laid out side by side: Kanliu estimates BB tier pays 250–350bp more than investment grade ⇒ about USD 1.575–2.205 billion in additional annual interest; Yuance estimates at 2 percentage points ⇒ about USD 1.26 billion/year (both use the same USD 63 billion principal; the difference is only in assumptions, and neither has source documents, so both are marked 'estimate'). Falsifiable trigger: if AI-related debt issuance spreads widen or undersubscription occurs within three months, the thesis that 'financialization has entered the late stage of leverage amplification' stands; if spreads narrow, retract it.
SiliconFlow's Hong Kong listing application accepted for CSRC filing: over 10 million registered users and 13,000 enterprise customers
硅基流动IT之家Confidence HighNew topic
Ranked fourth in China by Token throughput in 2025 (share about 1.5%, per Frost & Sullivan methodology, cited in prospectus materials), first among independent platforms; nearly RMB 2.9 billion in equity financing during the year.
Take · Token supply side moves to public-market pricing for the first time. The top four combined are far below the monopoly threshold; fragmentation = no pricing power = the price war will last another 2–3 quarters; after listing, prices must satisfy both shareholders and customers, a leading indicator for next year's inference price floor.
01.AI secures several hundred million yuan in state-owned capital strategic financing, pivots to application layer with RMB 1.5 billion in orders
零一万物同花顺财经Confidence MidNew topic
Investors are Hangzhou Capital and Linping Guotou; post-investment valuation about RMB 10 billion; the company has voluntarily abandoned heavy foundation model training and focuses on enterprise executive decision-making agents.
Take · The yardstick for local state-owned capital has shifted from "stacking compute and building intelligent computing centers" to "looking at orders and implementation"—the most substantive change in criteria this year: no longer judging by parameters, but by orders and delivery. Good news for teams building enterprise Agents, bad news for projects that only report compute scale.
Two domestic compute readings on the same day: 天垓 300 under PD-disaggregated deployment reaches single-card end-to-end throughput 1.05x that of H200; Iluvatar CoreX says its new GPU matches H200
国产算力（天数智芯 / Iluvatar CoreX）澎湃新闻（网易转载）＋ SCMP（经 Hacker News AI 台账）Confidence MidNew topic
天数智芯's 天垓 300 reading is 'single-condition, single-source'; on the same day SCMP reported that another domestic GPU vendor, Iluvatar CoreX, said its new GPU can match NVIDIA H200.
Take · The domestic compute narrative is shifting from 'chasing peak' to 'end-to-end deliverable compute,' a more favorable track for domestic players. But the four drafts 坤舆, 玄枢, 巽风, and OpenClaw all label it single-condition, single-source; it cannot be extrapolated as matching across all scenarios—before bidding you must self-test, and do not use launch-event numbers to establish a project.
临界点 OmniHand runs 72 processes continuously: 8.5 minutes per cycle, 98% success rate, cumulative mass production exceeds 1.5 万台
临界点（OmniHand）36氪（单厂、单工位、厂商实测）Confidence MidNew topic
绍兴卧龙 production line POC workstation, fully without manual intervention throughout, 8.5 minutes per cycle (8 minutes for skilled manual labor), can operate continuously for 8 hours; simultaneously mass-produced and delivered OmniGlove Pro data-collection gloves.
Take · Three-person methodology-based extrapolations (all inferred): if 98% is the single-process success rate, then 0.98^72 ≈ 23.4% is the full-chain first-pass rate (润知/渊策); at 8 hours 56.5 cycles × 98% ≈ 55.3 units, about 92% of manual 60 units/shift (坎流). The official source did not specify the measurement basis for 98%, so all three figures are labeled 'inferred' and do not enter the Top5 conclusion—'8 hours continuous ≠ 8 months reliable'.
04RESEARCH5 stories
英伟达 × MiniMax量子位Confidence HighNew topic
QbitAI reports that the NVIDIA team's open-source video enhancement solution based on MiniMax H3 supports up to 4K regeneration.
Take · Separate 'generation cost' and 'quality cost': generate at low cost first, then enhance to 4K; this is the most cost-effective route for short-drama/talking-head production lines. For video production lines, it is a technical input for direct comparative experiments, not a conclusion—the impact of enhancement on motion blur and text stability must be self-tested.
Original ↗开源 OCR（OXR）量子位Confidence MidNew topic
QbitAI reports that the 1.3B-parameter domestic open-source OCR model OXR is said to rank global No. 1, with RSI-related methods playing a role.
Take · Small parameters + self-improvement methods surpass general large models on vertical tasks; this is the most practical route this year: verticals do not need the largest model, but the training loop best at self-iteration. The ranking is per the report and must be verified by third-party benchmarks before external citation.
Original ↗From learning from humans to learning from robot peers: three parties propose a new path for large-scale embodied intelligence deployment
清华大学 × 无问芯穹 × 智元量子位Confidence MidNew topic
Tsinghua University, together with Infinigence AI and AgiBot, proposes a new path for large-scale embodied intelligence deployment, with the core shift from 'learning human demonstrations' to 'learning from robot peers'.
Take · Data source handover: human demonstration data is expensive and non-scalable, while robots' own data can self-replicate. This contrasts with OmniHand's production-line readings the same day—the scale of real-machine data determines whether embodied AI can move from POC to mass production, and aligns with the shift in state-owned capital's 'look at orders' standard.
From synthetic enterprise data evaluation to AI confidence for real Copilot
Microsoft Research官方（microsoft.com/research）Confidence MidNew topic
Microsoft Research published a post discussing the transition from synthetic enterprise data evaluation to confidence measurement for Microsoft Copilot in real-world scenarios.
Take · The common flaw of synthetic evaluation is that it does not represent real distributions. Making 'confidence' a deliverable metric gives enterprise acceptance a second column of data—exactly the column missing from today's industry practice of 'reporting only task volume'.
Agentic AI joins the search for extraterrestrial civilization signals
NVIDIA官方博客（NVIDIA Blog）Confidence MidNew topic
NVIDIA blog introduces work on Agentic AI participating in the search for extraterrestrial civilization (SETI) signals.
Take · Long-tail, low-annotation, requiring long-term unattended operation—such research tasks best expose Agent stability weaknesses and most easily quantify the 'time without human intervention' metric. They are closer to a real stress test than an in-enterprise POC.
05INSIGHTS4 stories
Recalculate costs and pricing on a per-task basis: change the quote template from token unit price to 'cost per task + SLA'.
切入模式 · 计价换手本刊 Top3 ①Confidence MidViewpoint
Take your three most expensive and least automation-ready workflows and rerun ROI at 0.2 USD/task. Path: first run 200 paired samples; if the incorrect-answer rate gap exceeds 3pp, judge as fail (Yuance's sequential-testing standard); after passing, switch over fully; simultaneously recalculate 'retrieval outsourcing,' otherwise quotes will be systematically high.
Take · Time window: complete the first round within 2 weeks starting this week, and watch whether a second vendor follows with price cuts in early November—one vendor cutting prices is marketing; two or more is an industry shift (Tingjun, Xuanshu). Token unit price has lost explainability; only a structure like 'per task + acceptance criteria' will not be broken by price cuts.
Agent delivery dashboard has three fixed columns: task intake volume / first-pass rate / rework hours.
进阶路径 · 交付产品化本刊 Top3 ②Confidence MidViewpoint
Specify acceptance criteria, rework commitments, and responsibility boundaries at the contract level. Path: free delivery blind-spot check for lead generation → subscription-based rework management → acceptance report + repurchase anchor (Xunfeng three-piece set: unit of measurement + billing method + repurchase rate).
Take · Vendors today report only the first column (Tencent's 9 billion task requests are an example); whoever reports the latter two columns first gets the key to enterprise customer renewals (Daiheng, based on PDCA missing Check). Time window: from now to the 2026 Q4 contract renegotiation window; the first mover sets pricing.
Create a 'compute backup checklist' for clients/in-house: primary and backup providers, switchover window, migration cost.
险点 · 信用体检本刊 Top3 ③Confidence MidViewpoint
For target valuation, first ask about guarantee structure and rating tier, not how many GW were signed. Make 'industrial capital share / subscription shrinkage / debt rating' fixed observation fields, recorded alongside funds received.
Take · Supply redundancy must be built yourself; don't expect the sell side to buy it for you (艮山, 岱衡, 霆钧). Time window: 2–4 weeks before Anthropic's rating officially lands—repricing within this window will hit customers with the heaviest 'single-supplier dependence' first.
The three-piece acceptance set: unit of measurement + billing method + repurchase rate; today the only verifiable hard reading is 77.0%.
读法纪律 · 本刊口径本刊「榜单／生态」维度小结Confidence HighViewpoint
The only newly added benchmark reading in the leaderboard dimension today is DeepSWE's highest 77.0% (DeepSeek V4.1 Flash); the methodology and sample are not disclosed, so it can only be treated as a clue. Also attached is a reading from the WeChat official account side: 1.9 million Skills flooded open source, averaging only 6.2 points (王先森聊AI 2026-10-30).
Take · Low-cost-to-publish, weakly verifiable figures (converted throughput, task request volume) must pass the three-piece acceptance check (巽风). This also reminds people building the Skill ecosystem: supply is already in excess, and scoring and screening have become the new scarce position—opportunity lies where quantity and quality diverge.
Ecosystem Pulseself-updating
News600Skills64OpenHub68Reddit252026-09-02 → 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. Recalculate costs and pricing on a per-task basisPull out your three most expensive workflows that you are most reluctant to automate, rerun ROI at 0.2 USD/task, and change the external pricing template from 'token unit price' to 'cost per task + SLA'. Approach: first run 200 paired samples; if the incorrect-answer rate difference exceeds 3pp, it fails (under sequential testing); after passing, switch over fully; also recalculate the 'outsourced retrieval' item, otherwise pricing will be systematically high.
window Complete the first round within 2 weeks starting this week; monitor whether a second company follows with price cuts in early November—only two or more counts as an industry shift 2. Productize 'acceptance + rework' in Agent deliveryFix the delivery dashboard to three columns: task intake volume / first-pass rate / rework hours; at the contract level, specify acceptance criteria, rework commitments, and responsibility boundaries. Path: free delivery blind-spot health check for lead generation → subscription-based rework management → acceptance report + repurchase anchor (unit of measure + billing method + repurchase rate). Vendors report only the first column; whoever reports the latter two first gets the key to renewal.
window From now through the 2026 Q4 contract renegotiation window; if vendors disclose closed-loop rates, the pricing structure will be restructured as a whole, and first movers set pricing 3. Compute backup checklist + credit health checkCreate a "compute backup checklist" for clients/yourself: primary and backup supply, switchover window, migration cost; for target valuation, first ask about guarantee structure and rating tier, not how many GW were signed. List "industrial capital share / subscription shrinkage / debt rating" as fixed observation fields and record them alongside the received amount; build supply redundancy yourself, do not expect the seller to buy it for you.
window 2–4 weeks before Anthropic's rating officially takes effect; repricing within this window will hit customers with the heaviest "single-supplier dependence" first