BunCosmos China AI Weekly

26 October to 2 November 2026

BunCosmos Research Lab | 2 November 2026 | buncosmos.ai

China's AI developments this week reached beyond model launches. Apple announced a timetable for bringing selected Qwen features to iPhones in China, Moonshot responded to problems serving its latest model, and DeepSeek shared research code for a more persistent coding agent. Infrastructure earnings and policy announcements also showed how deployment, operating reliability and access to technology are shaping the next stage of AI adoption.

Apple announces Qwen support for selected iPhone AI features

31 October | Product announcement

Apple said its first Apple Intelligence features for iPhone users in China will enter developer beta testing from 4 November and launch with iOS 27.3 in December. Its Alibaba partnership adds Qwen support to selected experiences in Siri, Writing Tools and Visual Intelligence. [1]

Apple's own generative models power Apple Intelligence. The announcement does not identify a specific Qwen checkpoint or establish that a downloadable Qwen model runs the whole system. It also excludes the next-generation Siri AI features announced at WWDC 2026. The partnership gives Qwen a new distribution channel through familiar iPhone applications, with the consumer rollout planned for December. [1]

Kimi responds to launch problems with a quota reset

26 October reporting | Follow-up to the previous weekly roundup

Reuters reported on 26 October that Moonshot AI apologised for rate limits and unnecessary quota consumption following the Kimi K3.1 Pro launch. The company offered a one-time reset equivalent to 25% of subscribers' monthly quota and said it was reducing unnecessary subagent calls and output length. [2]

This service response follows the launch covered in our previous China roundup. Moonshot's technical blog states that full model weights will be released by 4 November. For developers, the promised release is a separate milestone from the hosted access already available through Kimi products and its API. [3]

DeepSeek shares research code for a persistent coding agent

Early November coverage | Research prototype

DeepSeek's dsh-superagent prototype is designed to work on a single GitHub repository through DeepSeek Harness. It can receive requests through GitHub and optionally Feishu, carry out work in separate sessions, and respond in the original conversation, including with a pull request. A web panel shows items requiring human attention. [4]

The repository explicitly describes the code as a research prototype, not an official DeepSeek product or production-ready software. It also says the public implementation differs from the internal version and the accompanying design article; some proposed features remain future work. Its significance is the attempt to reduce the attention needed to manage coding agents, rather than proof that an autonomous engineering team is ready for deployment. [4]

Model evaluations and business

A revised benchmark exposes a larger gap for released weights

27 October | Artificial Analysis Intelligence Index v5

Artificial Analysis changed its composite evaluation, with 60% of the revised index based on private test sets or held-out answers. Its launch results distinguish models whose weights are promised from those already downloadable. [5]

Model	Score	Weight availability
Claude Opus 5.5 at max effort	58	Proprietary
Kimi K3.1 Pro	46	Not yet released in the index announcement
DeepSeek V4.1 Flash 1017 at max effort	32	Released
GLM-5.3 at max effort	32	Released

The 12-point gap includes Kimi's promised weights. Among models with released weights, the gap to the leader was 26 points. These are results under the revised test mix, not evidence that model capabilities suddenly deteriorated. A composite score also does not predict every user's task or deployment conditions. [5]

ZTE reports computing growth alongside pressure on profits

26 October | Unaudited third-quarter results

ZTE said computing revenue rose by more than 70% year on year in the first nine months of 2026 and reached 38% of total revenue. Its third-quarter revenue was RMB 38.81 billion, up 34.0%, while net profit attributable to shareholders rose 394.5% to RMB 1.31 billion. [6]

The wider reporting period gives a more mixed picture. The company's filing shows nine-month net profit of RMB 4.06 billion, down 23.71% year on year. ZTE also acknowledged pressure on gross margins and credited terminals and international markets alongside computing for revenue growth. The results demonstrate a growing computing business, but do not establish that AI alone caused the quarterly profit increase. [6, 7]

AI valuations feature prominently in the Hurun rich list

27 October | Hurun China Rich List 2026

Hurun ranked ByteDance founder Zhang Yiming first, estimating his wealth at RMB 670 billion, up RMB 200 billion from the previous year. Hurun attributed the increase to a higher company valuation supported by progress in AI. It also placed DeepSeek founder Liang Wenfeng seventh with an estimated RMB 280 billion. [8]

The list uses wealth estimates as of 1 September, although it was published on 27 October. It provides evidence of the financial importance attached to AI businesses and expectations. Personal wealth estimates and company valuations are not measures of realised AI revenue, profitability or technical performance. [8]

Policy and education

Export control consultation remains open until mid November

Ongoing | Proposal announced 16 October

China's Ministry of Commerce and Ministry of Science and Technology are consulting on additions to the technology export catalogue until 15 November. The proposal covers 40 technology items across AI, biomedicine and photovoltaics. Of these, 23 concern AI, including large models, agents, computing clusters and information processing. [9]

The proposed catalogue changes remain under consultation. Separately, the announcement already requires pre-export registration for covered transfers from 16 October until the revised catalogue takes effect. For AI, its stated scope excludes technology already in the public domain through open source or academic publication. The distinction between a proposed restriction and an existing registration requirement matters when describing the policy's current effect. [9]

China reports on international AI capacity building

26 October | Foreign Ministry briefing

At its 26 October briefing, China's Foreign Ministry discussed a progress report on its AI Capacity-Building Action Plan for Good and for All, released at the Global Governance Forum in Xiong'an. The spokesperson described cooperation on access to models, data and talent and emphasised developing countries' ability to build their own AI capabilities. [10]

The ministry presented access to technology and local development capacity as priorities for cooperation. This is China's stated policy position and its government's account of progress; practical outcomes will require separate assessment. [10]

Hong Kong plans AI learning within a new primary subject

30 October announcement | School rollout planned from September 2028

Hong Kong's Education Bureau published a provisional framework for an Information and Innovation Technology curriculum for primary schools, including a strand on the preliminary study of AI. SCMP reported that the subject will be introduced gradually from the 2028/29 school year, beginning with Primary One and Primary Four. [11, 12]

The reported framework includes simple interactions with AI using prompts or commands, with children as young as seven among the learners. The future timetable gives families and teachers an indication of how AI literacy will enter the new subject. [12]