AI Morning Brief · Oct 25, 2026
Top Stories
1.Follow-up (official launch): Moonshot's Kimi K3.1 Pro launches with deliberate cuts to its cyber-offensive capability
Reuters, Oct 24: Moonshot introduced Kimi K3.1 Pro while announcing it had stripped cyber-related data from the model's training corpus and reinforcement-learning environments, saying it aims "not to release open weights that would extend the absolute frontier of cyber capability," while retaining general coding skills; weights are due by Nov 4. The backdrop is China's draft AI export controls, published last week, that would ban AI-driven malware-generation technology and restrict AI-assisted vulnerability-discovery tools (open for comment until Nov 15). This upgrades yesterday's "official teaser + press-reported launch" to a formal launch with an official safety stance; only the weights remain pending (Oct 24 / Reuters)
Sources: reuters.com · unite.ai
2.Backfill: UK announces a £100M AI Biosecurity Lab, with OpenAI and Anthropic offering up to £200M in model credits
the UK government announced Oct 23 the world's first dedicated AI biosecurity organisation (AIBL) with £100M of government funding; OpenAI is contributing £100M and Anthropic up to £100M in model/compute credits — credits, not cash (per the BBC's framing); Google DeepMind and Basecamp Research are research partners, and the lab sits inside a roughly £600M AI safety and cyber package. The official page was updated Oct 24 with follow-up coverage unpacking the figures; note that sources differ on the package total (£600M vs $800M), reflecting different conversions and counting (Oct 23–24 / Tech Insider, Thames Compass)
Sources: tech-insider.org · thamescompass.com
3.Backfill: SpaceX's Super Intelligence Partner Network signs 130+ companies as Grok Bot passes 800,000 weekly paid users
Bloomberg reported Oct 22–23 that SpaceX launched a partner network for companies deploying or building on Grok, with Apollo, NVIDIA, Snowflake, Databricks, Bain and Booz Allen among 130+ early joiners; some partners help enterprises roll out Grok, others build plugins, and members can take SpaceX's Forward Deployed Engineer course. Context: on the Oct 20 earnings call, president Gwynne Shotwell said Grok Bot had over 800,000 weekly paying users in October, double the prior month; SpaceX's AI unit turned a Q3 profit largely from renting Memphis data-centre compute rather than Grok subscriptions. Missed by the last edition (Oct 22–23 / Bloomberg relay)
Sources: allusanewshub.com · finance.biggo.com
4.Follow-up: DeepSeek and MiniMax win approval for global AI deployment; Oct 24 analysis clarifies APIs and open source fall outside the controls
Reuters, Oct 23: CCTV-affiliated Yuyuan Tantian disclosed that DeepSeek and MiniMax completed registration under China's draft AI technology-export rules, clearing open-source release and overseas licensing; MiniMax confirmed on Oct 23. Oct 24 follow-up coverage clarified that API/web/app access and already-public open-source releases do not count as regulated exports, while the draft's 23 AI categories remain open for comment until Nov 15. A backfill plus a same-window interpretive follow-up (Oct 23–24 / Reuters, AI Affairs)
Sources: reuters.com · aiaffairs.com
5.Backfill: HKEX launches an AI Index and a Semiconductor Index spanning Alibaba, Tencent, Zhipu and MiniMax
HKEX launched Oct 23 the HKEX AI Index (40 stocks across five layers: chip design, data centres, AI infrastructure, models, software/applications) and the HKEX Semiconductor Index (30 stocks including ASMPT, SMIC and Hua Hong), licensing China Southern Asset Management to issue a mainland ETF tracking the AI Index; southbound flows on Oct 23 showed net buys in Zhipu (HK$1.739B) and Tencent (HK$1.296B) against net sells in Alibaba and SMIC. Surfaced in a BigGo summary published 09:25 Beijing time Oct 24 (Oct 23–24 / BigGo)
Sources: finance.biggo.com
6.Cambricon Q3: revenue RMB 3.531B (+104.5%) and profit RMB 1.404B (+147.7%), but below analyst expectations
Cambricon reported Oct 23 Q3 revenue of RMB 3.531 billion, up 104.47% year on year, and net profit of RMB 1.404 billion, up 147.74%, with nine-month revenue of RMB 9.526 billion; analysis published Oct 24 notes revenue fell well short of a roughly RMB 6.26 billion analyst expectation, attributed to manufacturing and HBM supply constraints — the expectation figure and attribution are relayed via Bloomberg/SCMP and are not verified directly against the filing (see Unverified). Profit surge and supply limits are two sides of the same story (Oct 23–24 / KuCoin flash, AI Weekly)
Sources: kucoin.com · aiweekly.co
7.Amazon CEO: AWS capacity is largely sold out through 2027, with 2026 capex at about $220B
relayed from Reuters (Oct 24): Amazon CEO Andy Jassy said Friday that AWS capacity is "largely sold out for 2027, and a good portion of 2028," on strong demand from AI labs and enterprises; Amazon's 2026 capex guidance has been raised by $20B to about $220B. A major cloud supply-demand signal (Oct 24 / Reuters relay)
Sources: catenaa.com
8.CNBC: NVIDIA's role as an AI investor faces its first big test as neocloud stocks slide
CNBC reported Oct 24 (via relays; the original CNBC link was not directly obtained, see Unverified): NVIDIA is no longer just selling chips but financing the neocloud companies that buy them through investments, credit and prepayments; after several neocloud stocks fell this week, markets question whether debt-driven AI infrastructure expansion can generate enough return, leaving NVIDIA exposed to customer credit risk. A market-narrative turn, not a product fact (Oct 24 / CNBC relay)
Sources: vespernews.com · albertpham.info
9.Backfill: Sweden's Government Offices will block ChatGPT and Claude from Oct 26, leaving Copilot as the only approved tool
Sveriges Radio's Ekot reported Oct 23 that Sweden's Government Offices (Regeringskansliet) will block ChatGPT and Claude on staff service computers from Monday Oct 26, steering employees to Microsoft Copilot, which passed an internal security review; digitalisation director Mikael Wettercrantz warned that AI services used without a contract carry information-security consequences; the Tax Agency and the National Board of Institutional Care already have similar restrictions, and a Transport Administration audit found staff using 10 open AI services with sensitive data entered. A notable government AI-governance case (Oct 23 / CryptoBriefing, LavX)
Sources: cryptobriefing.com · news.lavx.hu
Platform & Model Tracking
OpenAI (GPT/ChatGPT)
the $30B raise is still negotiating, not closed (Oct 24 relayed terms unchanged: up to $30B at a $1.25T pre-money valuation, UAE investors and Coatue leading); a participant in the UK AI Biosecurity Lab (see Top Story 2, model credits not cash); no new flagship model release in the window — Forkast's Oct 24 check notes the supposed Oct 28–31 "release pegs" for three labs are tracker predictions, not vendor announcements; GPT-6.1 "Astra" remains cancelled (last edition's correction stands) and GPT-6.1 Terra remains a paywalled rumor
Sources: neoteo.com · forkast.news
Anthropic (Claude)
still no public S-1 — an Oct 24 SEC EDGAR search found no match (only the confidential draft submitted Jun 1 is company-confirmed; a late-October/November public filing is press reporting); a participant in the UK AI Biosecurity Lab (see Top Story 2, up to £100M in model credits); Forkast's check: the newsroom's latest items are still the two Oct 22 posts (Agent Skills, life sciences) and the latest model announcement remains Oct 7's Claude Haiku 5.5; Claude formal-proof claims and an Oct 24 outage are single-source (see Unverified)
Sources: neoteo.com · tech-insider.org
Google DeepMind (Gemini)
minor launch: ChromeOS 156 (Oct 24) brings Gemini media understanding to Chromebook Plus and Gemini features in Chrome file and quiz flows; Google Messages is pushing a Gemini scam-detection "Enhanced Protection" beta (Oct 24, single-source relay, no Google announcement seen — see Unverified); Agent Compute Engine was announced Oct 22 (missed by earlier editions, backfill note): one hardware-isolated microVM per agent with zero credentials in the sandbox, customer preview in the coming months; Gemini 4 Carbon still has no official release
Sources: ghacks.net · techtimes.com
Meta
no new Llama or Muse model, pricing or availability change in the window; the Japan launch (Oct 23) and Class C stock (Oct 21) were covered last edition
xAI (Grok)
qualifies: the Super Intelligence Partner Network (see Top Story 3, backfill); Grok Bot's 800,000+ weekly paid users (Oct 20 earnings call, backfill figure); Grok 5 still has no official release or date (flagship remains Sept 21's Grok 4.7); the Virginia AG litigation risk remains a single MLex source from the last edition
Sources: allusanewshub.com · teslanewswire.com
Microsoft
no major updates in the window; Copilot is the sole approved tool in Sweden's government block (see Top Story 9), a procurement and compliance outcome rather than a product launch
NVIDIA
the neocloud financing-risk story (see Top Story 8, CNBC market analysis, not a product fact); the substrate-capacity plan (2.5× in three years, a $1.1B prepayment to KYEC) was announced Oct 21 and reported Oct 24, with the prepayment detail single-source (see Unverified); no new chip or model launch
Sources: defenseworld.net · bytevyte.com
Mistral
Large 4 weights still not released as of Oct 24 (official commitment is end of October; a VentureBeat-reported Oct 27 date is single-source, not a Mistral commitment); Forkast's Oct 24 check: the Oct 28–31 "release pegs" are unannounced tracker predictions
Sources: beam.ai · forkast.news
China AI Companies
Moonshot AI (Kimi)
qualifies: K3.1 Pro officially launched with deliberate cyber-capability cuts (see Top Story 1 — Reuters, Oct 24); spec recap: 3-trillion-parameter-class open-weight model, 1M context, API at $2 input / $10 output per million tokens (Highspeed roughly double); weights due before Nov 4; exploit-benchmark figures (81.0% vs K3's 33.3% and DeepSeek-V4.1-Flash's 85.7%) are single-source (see Unverified); K2.8 remains Preview
Sources: reuters.com · nogentech.org
DeepSeek
qualifies: global-deployment registration approved (see Top Story 4 — Reuters, Oct 23, backfill, with Oct 24 interpretive follow-up that APIs and open source fall outside the controls); financing commentary on Oct 24 evening Beijing time rehashed the ~RMB 100B round (CATL and Tencent as largest backers, ~RMB 500B pre-money, ~$1B annualised revenue) with no new company confirmation or investor list; V4.1-Flash-1017 remains the latest model; DeepInfer was an Oct 20 release (outside the window)
Sources: reuters.com · aiaffairs.com
MiniMax
qualifies: global-deployment registration approved (see Top Story 4, self-confirmed Oct 23); M3.1 Flash weights were open-sourced Oct 23 (last edition); no new product facts in the window
Sources: reuters.com
Tencent (Hunyuan/Yuanbao)
qualifies (backfill): Hunyuan Hy3 open-sourced Oct 23 under Apache 2.0 for free commercial use — 295B MoE, 21B activated, 256K context, with agent and coding capability claimed +30% over the preview; weights on Hugging Face, ModelScope and GitHub; no new release in the window
Sources: news.metal.com
Alibaba (Qwen)
no new model or product launch in the window; Qwen4 still has no official announcement or date (in training since the Sept 22 announcement; anonymous Arena testing is an Oct 23 community-leak claim, single source); minor ecosystem note (auto-generated listings): GGUF conversions of Qwen2.5-3B-Instruct and Qwen2.5-Coder-3B appeared Oct 24 on Huawei's Modelers.cn for local inference in 2–3GB of memory — not new Alibaba models, only community repackaging
Sources: orcarouter.ai · aichina.news
ByteDance (Doubao)
no new product facts in the window (the Doubao blocklist expansion from 151 to 166 apps was Oct 22, outside the window)
Zhipu (GLM)
no new model or pricing release in the window (the GLM price hike was Oct 22); southbound net buys in Zhipu reached HK$1.739B on Oct 23 (see Top Story 5)
Baidu
no new product facts in the window
Kuaishou (Kling)
a third-party claim of first integration with the Kling 4.0 API is vendor press-release material (SelfyzAI, Oct 23), not independently verified and not counted as a product fact; the Kling 4.0 API itself launched Oct 22
Sources: eurozone.einnews.com
Manus / StepFun / SenseTime
no substantive new developments in the window
Policy and markets
export-registration approvals (see Top Story 4); the UK biosecurity lab (see Top Story 2, international policy context); the HKEX AI and Semiconductor indexes (see Top Story 5); Cambricon Q3 (see Top Story 6); Q3 data-centre tenders from China's three carriers — 225+ tenders totalling RMB 36.588B (Mobile RMB 24.57B, 67.1%; Telecom RMB 7.64B; Unicom RMB 4.38B), roughly 90% in the eight national computing hubs — a single weekly synthesis (see Unverified), signalling capex migration from telecom networks to AI compute
Sources: kimkj.com
Worth Reading
1."Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks" — Reuters, 2026-10-24
the full K3.1 Pro launch and deliberate cyber-capability cut — the core tension of the open-weight era, where stronger models are harder to govern, and the first major vendor response under China's draft export controls
Sources: reuters.com
2."Three Frontier Release Pegs Packed Into Oct 28-31. None of the Vendors Has Announced Them." — Forkast, 2026-10-24
a newsroom-by-newsroom check separating tracker predictions from vendor announcements — a cold shower for the late-October release-rumor wave
Sources: forkast.news
3."UK launches AI Biosecurity Lab with £100m of government funding" — Thames Compass, 2026-10-24
untangles the three funding layers — £100M government cash, up to £200M in model credits, a roughly £600M package — so credits are not mistaken for cash
Sources: thamescompass.com
Unverified Items
1.
White House pre-release review of new models from Google, Microsoft and xAI: a Forbes mirror page (pages.dev) whose URL path contains 2026/05/05, inconsistent with the Oct 24 timestamp shown in search — the original text could not be directly verified, so it is not in the main text.
2.
Claude formal proofs (Claude + Theorem generating ~180M lines of Rocq proof code across 123 OSS-Fuzz libraries): TokenPost, Oct 24, single source.
3.
Claude outage on Oct 24 (05:17–05:42 UTC, about 24 minutes): nocrash.io outage tracker, single source.
4.
Google Messages Gemini scam-detection beta: single-source relay citing Digital Trends and 9to5Google, with no Google announcement seen for the day.
5.
NVIDIA's $1.1B prepayment to KYEC for capacity reservation: single Oct 24 briefing page, no independent announcement seen.
6.
Kimi K3.1 Pro post-curb exploit-test scores (81.0% vs K3's 33.3% and DeepSeek-V4.1-Flash's 85.7%): Nogentech secondary analysis, no Moonshot chart seen.
7.
Cambricon's "below a roughly RMB 6.26B analyst expectation" and the HBM supply-constraint attribution: Bloomberg/SCMP relayed via AI Weekly, not verified against the filing directly.
8.
Three carriers' Q3 data-centre tenders (225+ tenders, RMB 36.588B): single weekly synthesis, not a single-day announcement.
9.
CNBC's neocloud financing-risk report: relayed summaries only; the original CNBC link was not directly obtained.
10.
Standing open questions with no new progress in the window: Grok 5, Gemini 4 Carbon, Qwen4, Mistral Large 4 weights (due end of October), Anthropic's public S-1, OpenAI's $30B financing; Kimi K3.1 Pro weights due before Nov 4; the Zhipu "3.5-trillion-parameter" rumor and the final details of DeepSeek's round (the company remains "no comment") have no new sources.