AI Morning Brief · Oct 16, 2026

Coverage window: 2026-10-15 07:00 → 2026-10-16 07:00 (Beijing Time)

Research note: English edition written primarily from English-language sources; some items cite Chinese-language sources.

Top Stories

1.Meta launches open-weight Muse Spark Open (816B/42B active), jointly optimized with Apple for M5 Ultra

Meta's strongest open-weight model to date: 816B total / 42B active MoE, 1M-token context, native multimodal, built for coding and agent workloads; free weights via Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry; Meta and Apple jointly announced deep optimization for M5 Ultra/MLX, running locally on two M5 Ultra Mac Studios (2026-10-15, TechCrunch/Reuters)

2.NVIDIA ships Nemotron 3.5 family: Super VL 120B + Bolt 74B-A9B now live, open weights on Hugging Face

Super VL: 120B params/12B active, 1M context, multimodal, for data-center resident agents; Bolt 74B-A9B: 9B active, native 2-bit quantization fits 24GB VRAM, 256K context, available today; smaller Bolt E4B (4GB) due 10-23; OpenMDW-1.1 license (commercial use allowed) (2026-10-15, multiple outlets)

3.Anthropic discloses Q3 revenue of $16.8B (+45% QoQ), pushes IPO past November midterms

Revealed at an all-day investor meeting on 10-14: Q3 $16.8B (Q1 $4.8B, Q2 $11.6B), second consecutive quarter of adjusted profitability with positive free cash flow; $2T+ target valuation; SemiAnalysis cut its annualized revenue forecast to ~$74B (previously $100B+) (2026-10-15, FT/Reuters/Bloomberg/PYMNTS)

4.OpenAI launches Rosalind Academic Plan: up to 30 free seats for US university labs to use GPT-Rosalind life-sciences model

Eligible labs get up to 30 Business Standard seats free for 12 months with access to GPT-Rosalind (frontier reasoning + Rosalind workbench for medicinal chemistry, genomics, wet-lab assistance) via trusted access; research data not used for training by default; no API yet (2026-10-15, OpenAI official blog)

5.OpenAI publishes API deprecation timeline: GPT-4o family sunsets 2027-01-14, older models cut 2026-10-23

GPT-4o/4o mini/4.1/4.1 mini leave the API on 2027-01-14 (recommended replacement gpt-6-luna); gpt-4-turbo, o3-mini, o4-mini, gpt-4.1-nano shut down on 2026-10-23 (notified 10-14, shellypalmer)

6."Follow-up" Tim Cook confirms in Shanghai: China Apple Intelligence approved, Alibaba Qwen models shipping in late-October beta

Apple Intelligence plus Alibaba Qwen large models approved; test version launches late October with real-time translation, writing tools, photo cleanup; Siri AI not included this time; Mark Gurman: full China launch via iOS 27.3 in December with Qwen and Wanxiang (2026-10-15, Sina Finance/ITHome; Chinese-language sources)

7.Anthropic takes Claude Code's Code Review to general availability (10-15)

Out of research preview for Team and Enterprise orgs; ~3x faster reviews (20 min to 6-10 min), new Quick/Standard/Deep review depths; eligible Enterprise orgs can claim at least $10,000 in promo credits before 2027-01-15 (official blog + Unite.AI)

8.DeepSeek suffers major service outage on 10-15 morning; company says cause located, fix underway

Web and app returned "server temporarily unavailable"; DeepSeek said the incident affected chat, the DeepSeek V4.1 Flash API and another component; StatusGator logged 86 outage reports in 24h across Luxembourg, Poland, New York and more, lasting over an hour (single source, Techritual Hong Kong; Chinese-language source)

9.AI infrastructure funding streak: Modal raises $750M Series D at $15.75B, Labelbox $100M+ Series E, 64MW Mumbai AI data center breaks ground

Modal: $750M Series D led by Accel, valuation tripled from $4.65B (May C round) to $15.75B in five months, $1B+ annualized revenue (single source, CEO blog via PYMNTS); Labelbox: $100M+ Series E (equity + credit, Infinitum Partners lead) for reinforcement learning and enterprise agents (single source); Colt DCS x RMZ break ground on 64MW AI-ready Mumbai 3 data center, fully pre-leased (single source); Fermi Inc. discloses 15-year 222MW/$6.5B lease with TensorWave's Matador project (single source)

10."Follow-up" OpenAI–Apple trade-secret trial: judge says injunction "highly likely", OpenAI concedes

Apple sued OpenAI in July over poached employees allegedly taking unreleased-product IP; Judge Edward Davila signaled a likely injunction; OpenAI agreed not to use the secrets and will submit forensic images of 14 employees' devices (single source, Bloomberg via HinduBusinessLine)

Platform & Model Tracking

OpenAI (GPT/ChatGPT)

Launched Rosalind Academic Plan (up to 30 free Business Standard seats for 12 months for eligible US university labs to use the GPT-Rosalind life-sciences model via trusted access; research data not used for training by default); published API deprecation timeline (GPT-4o family sunsets 2027-01-14, gpt-4-turbo/o3-mini/o4-mini/gpt-4.1-nano shut down 2026-10-23, replacement gpt-6-luna); Europe report claims AI agents now account for 50%+ of EU enterprise platform token output (single source, ANI). Trade-secret trial vs Apple is non-product news.

Anthropic (Claude)

Claude Code's Code Review went generally available on 10-15 (Team/Enterprise, ~3x faster, three review depths, $10,000+ promo credits for eligible Enterprise orgs); Q3 revenue $16.8B (+45% QoQ), IPO pushed past November midterms; FIS's AML investigation agent built with Anthropic is now commercially available to US banks (single source, PYMNTS citing company release).

Google DeepMind (Gemini)

No significant updates. Minor single-source item: third-party router ZenMux now offers Gemini 4 Argon via its API (not an official Google move).

Meta

Released open-weight Muse Spark Open (816B/42B active MoE, 1M context, native multimodal, free weights on HF/SageMaker/GCP Model Garden/Foundry) plus rare partnership with Apple optimizing it for M5 Ultra/MLX (local runs on two M5 Ultra Mac Studios); @AIatMeta claimed Muse Wonder benchmarks beat GLM 5.3 and Kimi K3 (performance claim, unverified, not a product change).

xAI (Grok)

No significant updates. Grok 5 not released; Grok 4.7 remains the latest general model.

NVIDIA

Shipped Nemotron 3.5 family (Super VL 120B/12B active/1M context; Bolt 74B-A9B/9B active/24GB VRAM/256K context live; Bolt E4B 4GB due 10-23; weights on Hugging Face under OpenMDW-1.1); Multiverse Computing's Comet 60B compressed Super VL variant announced (single source); GeForce NOW adds DLSS 4.5 controls; The Information reports NVIDIA plans an equity investment in d-Matrix (single source); Huang told Bloomberg NVIDIA would consider cutting chip supply to customers developing "unsafe AI" (single source).

Mistral

No significant updates.

China AI Companies

DeepSeek

Major service outage on 10-15 morning (web, app, V4.1 Flash API affected; cause located, fix underway; single source). Follow-up: QbitAI testing found the deepseek-flash API noticeably stronger since 10-12 (possibly v4.2-flash or a new checkpoint); official statement still absent (single source). No new financing developments. Third-party: OpenCode now offers DeepSeek V4.1 Flash free (200K context); Prime Intellect published a bf16 re-export of DeepSeek-V4.1-Flash on Hugging Face.

Alibaba (Qwen)

Tim Cook confirmed China Apple Intelligence approval with Alibaba Qwen models shipping in a late-October beta (major partnership milestone; Chinese-language sources).

ByteDance (Doubao)

"Follow-up" SAEP refusal list updated to 155 apps (66 Tencent, 23 Alibaba, 14 Kuaishou, 11 Meituan; developers may file refusal statements at any time); companion-agent data retention: read-only window for decommissioned companion agent configs/chats ended 10-15 (single source); Douyin's Double-11 (launched 10-15) distributes coupons across Doubao, Hongguo short dramas and Toutiao, making Doubao a coupon entry point (The Paper; Chinese-language sources).

Tencent (Hunyuan/Yuanbao)

No significant updates; "Yuanbao Assistant" standalone-app rumor still has no new sources.

Moonshot (Kimi)

K3.1 confirmed still unreleased — checked against EvalMap's release list, Moonshot's official model list/platform changelog, models.dev, OpenRouter and Vercel directories; no K3.1 entries anywhere.

Zhipu (GLM)

No significant updates (no new GLM-5.3/GLM-Legal facts; only a Japanese media deep-dive, no new facts). Personnel: Yuki Chen (former head of 360's vulnerability research institute) joins Zhipu (The Paper; Chinese-language source).

Baidu

Baidu World 2026 set for November 12 (announced 10-15): "Next Act, Intelligent Organizations" theme, Robin Li keynote, "chip-cloud-model-agent" full-stack AI upgrades, 12 sub-forums, 30+ AI open classes, 6000-sqm AI expo (official WeChat account; Chinese-language source).

MiniMax

No significant updates.

Manus

"Follow-up" OPPO x Manus details (10-14): Agent capabilities integrated into ColorOS 17's AI Mind Pilot, debuting on the Find X10 series; on-device tasks are not shared with Manus; first system-level smartphone partnership since Manus regained independence after Meta's acquisition (BlockBeats + Owlting; Chinese-language sources).

Policy & funding

Microsoft retires Outlook's "Themes by Copilot" (and the "Wrap up your day" inbox summary) globally in mid-to-late October per an official admin notice, as Outlook Copilot pivots to a simpler chat design (multiple outlets citing Microsoft's notice).

Worth Reading

1."Anthropic quarterly revenue jumps 45% before planned IPO" — PYMNTS, 2026-10-15

unpacks the $16.8B Q3 disclosure, two straight quarters of adjusted profitability, the $2T+ target valuation vs SemiAnalysis's $74B annualized forecast, and the post-midterms IPO timing.

2."Block 披露与 OpenAI 联合打造的 AI 安全体系 Daybreak" — tech-insider, 2026-10-15

Block's security team built an automated attack-chain replay system on OpenAI's "Defense Factory" — an AI agent chaining multi-stage simulated attacks from SQL injection to privilege escalation to lateral movement in about 2 hours, with humans retaining final sign-off on every deployed fix. Single source, unverified.

3."Users keep rejecting Copilot and now Microsoft is killing its AI themes in Outlook" — Windows Latest, 2026-10-13

cites Microsoft's official admin notice retiring "Themes by Copilot" globally in mid-to-late October (alongside the "Wrap up your day" summary), as Outlook Copilot pivots to a simpler chat design.

Unverified Items

1.

Jensen Huang's "consider cutting chip supply for unsafe AI" remarks: Bloomberg original via a re-report chain (cryptofox/panews/heygotrade), single-source chain.

2.

NVIDIA's planned equity investment in d-Matrix: The Information original via embers.news, single-source chain.

3.

Multiverse Computing's Comet 60B (compressed Nemotron 3.5 Super VL): GlobeNewswire press-release single source, unverified.

4.

deepseek-flash "getting stronger": QbitAI testing via 投资界 reprint, no official statement, unverified.

5.

FIS x Anthropic AML investigation agent: PYMNTS citing company release, single source, unverified.

6.

Labelbox $100M+ Series E: PR Newswire via Morningstar, single source, unverified.

7.

Bloomberg Law's Microsoft Copilot connector (MCP): PR Newswire company announcement, single source, unverified.

8.

Claude Opus 6 / OpenAI "Bell" rumors: tech-insider.org debunked on 2026-10-15 — no official word from either company; Anthropic's current flagship remains Opus 5.5 (released 9-22). Rumor dismissed.

9.

Continuing from yesterday's unverified list (xAI renaming to "SpaceX Super Intelligence", Gemini 4 "Carbon", Tencent "Yuanbao Assistant", Moonshot K3.1 release, Nomura's 5–6 trillion-parameter claim, Sarvam AI funding, Astribot funding, Microsoft Maia, Grok 5, NVIDIA–Reflection AI talks, Infinigence IPO filing, ByteDance world model, deepseek-flash single source [→ follow-up with real testing in today's report], Gemini 4 Argon single source [→ now multi-source landed], Anthropic IPO roadshow [→ now formal disclosure]): no new sources in this window, not repeated.