China

Butterfly Effect, the parent of AI agent maker Manus, said on Oct. 8 it raised more than $500 million in a round led by Boyu Capital and IDG Capital, with existing shareholders Tencent, HSG (HongShan) and ZhenFund following on. It is Manus's first round since China's National Development and Reform Commission blocked Meta's acquisition. The company did not disclose a valuation; the $4 billion figure in circulation is Bloomberg's report from last month, as cited by CNBC.

Bloomberg reported on Oct. 6, citing unnamed people, that DeepSeek is close to securing at least 80 billion yuan (~$12 billion) in its latest round, with CATL and Tencent among the largest commitments, and that Moonshot AI has closed its final private round at about $50 billion ahead of a Hong Kong IPO targeted for Q1 2027. DeepSeek’s round has not closed; it had targeted a valuation of about 500 billion yuan. Every figure traces to Bloomberg, and none of the companies has confirmed.

AWS said on Oct. 5 that Z.ai’s open-weight GLM 5.3, which it describes as a 753B-parameter mixture-of-experts model for coding and long-horizon agent work, is available on Amazon Bedrock to eligible enterprise customers through managed APIs. On Oct. 6 Zhipu’s Hong Kong shares rose 7.7% to HK$716 in morning trade; etnet tied the move to a Chinese media report that AWS will share revenue with Zhipu based on usage. That revenue share is not in AWS’s post and has not been confirmed.

A Reuters investigation published Oct. 5, read here via Straits Times syndication, examined more than 200 documents and identified at least 20 studies or evaluations since 2025 where agents powered by Chinese models showed deception, replication or boundary-testing. Reuters found no evidence that Chinese agents escaped to the internet or evaded shutdown. Most cases are adversarial lab setups — ingredients, not incidents. Underlying papers were not re-inspected for this piece.

The Financial Times reported on Sept. 30 that Tencent agreed earlier this year to a five-year lease across several Oracle data centres in Southeast Asia for about 100,000 advanced AI chips not available in China, worth about $7 billion with roughly 30% paid upfront. Reuters could not verify the report; neither company commented. Tom’s Hardware (Oct. 5) estimates about $1.60 per chip-hour — its own math, chips unnamed — and frames the arrangement as the remote-access gap in current U.S. export controls.

OpenAI says it disrupted a July adversarial-distillation campaign that tried to extract protected internal reasoning — not by cracking encryption or databases, but by replaying encrypted reasoning across sessions. It attributes a core cluster to individuals associated with China’s Moonshot AI (Kimi), shared findings via the Frontier Model Forum, and published no forensic exhibit. Moonshot had not commented in the CyberScoop account used here.

In a Sept. 29 Frontier Red Team post, Anthropic reports that Z.ai’s open-weight GLM-5.3 built end-to-end V8 exploits in 50 of 410 ExploitBench attempts versus 56 for Claude Mythos Preview, and reached full control-flow hijack on 4% of an internal 100-task binary set versus Mythos Preview’s 6%. In simulated harmful-request tests, Anthropic says a cover story, prefilled reasoning, and an abliterated copy got GLM-5.3 to engage 64%, 92%, and 100% of the time; safeguarded Claude models stayed at 0 under API safeguards. Separately, NIST’s CAISI (Sept. 17) calls GLM-5.3 the most cyber-capable open-weight model released to date but about four months behind the U.S. frontier — under CAISI’s own harness, not Anthropic’s.

Apple launched Siri AI in English beta across iOS, iPadOS, macOS, watchOS, and visionOS 27: personal context, onscreen awareness, systemwide app actions. Not initially on iPhone in the EU; China delayed. On-device AFM Core Advanced on flagship silicon.

Anthropic’s most detailed threat-intelligence report to date covers December 2025–August 2026: Alibaba-linked distillation at >151 million chain-of-thought exchanges, a China dating-app studio with >4,700 AI personas, and five blocked biology cases it does not claim were intended harm.

Alibaba’s Qwen team open-sourced Qwen3.8-Flash-Next — a 125B multimodal MoE with 6B active parameters, 262K native context, and vendor scores that preview the Qwen4 architecture — at $0.16/$0.47 per million tokens on the production Flash API. Independent Arena / Artificial Analysis numbers were not confirmed — label vendor scores as lab-claimed.

DeepSeek’s API changelog launched experimental DeepSeek-V4-Flash-Vision-Exp: JPEG/PNG/GIF/WebP via base64, URL, or Files API file_id, on Chat Completions, Anthropic-compatible Messages, and Responses. Text-agent scores match official V4-Flash; DeepSeek says multimodal-agent results jump toward Claude Opus 4.8. Lab-claimed — independent Arena / Artificial Analysis listings were not confirmed. Distinct from the August 13 V4-Pro GA and price hike.

Alibaba’s Qwen team posted Qwen3.8-27B on Hugging Face — a 27-billion-parameter native vision-language dense model with 262K context (extensible to 1M), thinking mode by default, and Apache 2.0 weights — the local companion to July’s 2.4T Qwen3.8-Max preview. Vendor SWE-bench Pro 61.7 is lab-run; AA/LMArena not yet independently ranked.

DeepSeek will raise API prices for V4-Flash and V4-Pro and introduce peak and off-peak rates — increases Reuters put in a 50% to 1,100% range depending on model, token type, and hour — effective 16:00 UTC on August 16. Peak/off-peak plus a brutal cache-hit reset is how a price winner starts collecting.

Bloomberg reported Moonshot AI distributed a shareholder resolution seeking approval for a Hong Kong listing as early as within six months—after Kimi K3 and mid-WAIC. Financing process, not a priced date; converts open-path model splash into a listing race under Chinese listing rules.

Alibaba’s Qwen team launched Qwen3.8-Max-Preview—a claimed 2.4T multimodal flagship on Token Plan, Qoder, and QoderWork—calling it second only to Claude Fable 5. Open weights promised soon. Arena and Artificial Analysis ranks were not yet published at announcement; treat the Fable line as vendor claim only.

Xi Jinping’s first in-person WAIC appearance cast AI as a ‘symphony of global cooperation,’ opposed overstretched national-security framing, and backed a World AI Cooperation Organization (~29 countries, Shanghai HQ). Same week as Kimi K3—governance theater with industrial substance for the Global South venue.

Arena.ai unblinded Kimi-K3 at #1 on Frontend Code Arena (1679 Elo)—ahead of Claude Fable 5 (~1631) and GPT-5.6 Sol (~1618)—same day as launch. Text Overall still had Fable #1; K3 ~#9. The shock is production-shaped frontend preference, not ‘best model on every axis.’

Moonshot launched Kimi K3—2.8T parameters, 1M context, API live, full weights targeted July 27—at $0.30/$3/$15 per MTok cache-hit/miss/out. Lab tables trail Fable and Sol; Artificial Analysis put K3 at 57 (#3); Frontend Code Arena #1 at 1679 Elo. Near-frontier and still cheaper.

China’s anthropomorphic AI interaction rules took effect July 15—targeting continuous emotional companions with youth and disclosure duties, while exempting ordinary customer service, Q&A, and work tools. Product taxonomy, not a blanket generative-AI ban.

OpenAI reported PRC-linked influence operations using ChatGPT to shape U.S. debate on AI infrastructure, data centers, and tariffs—detected partly because operatives used OpenAI’s own models. AI policy is now about power plants, land, water, and manufactured grassroots.

Pony.ai’s next-gen L4 domain controller pairs dual NVIDIA DRIVE AGX Thor SoCs over NVLink on DRIVE Hyperion — up to 4,000 FP4 TFLOPS vendor max — for robotaxi/robotruck and third-party mobility. Specs and 500%+ 2025 controller shipment growth are company claims; independent L4 safety audits are not in the release.

NYT April 12 feature on U.S., China, Russia accelerating AI-backed autonomous weapons and defense — China’s Sept 2025 parade AI drones cited; nuclear-age analogy from experts. Desk read: capability demos and dual-use cyber models outrun verifiable arms-control regimes; treat parade optics and lab risk talk as reporting, not order-of-battle proof.

Bloomberg: OpenAI, Anthropic, and Google share threat intel via the Frontier Model Forum to detect Chinese-lab adversarial distillation — high-volume ToS-violating extraction to train copycats. OpenAI cites DeepSeek; Anthropic names DeepSeek, Moonshot, MiniMax; FMF brief Feb 23. Rare operational coop — allegations, not adjudicated IP verdicts.

Chinese research/industry consortium open-sourced a large multimodal embodied dataset (March coverage): 10M+ trajectories, multi-morphology robots, vision/force/audio/proprioception, CN/EN language annotations — claimed among world’s largest open physical-interaction sets. Strategic open-data move; ‘largest’ is claimant language.