DeepSeek 传融资约 150 亿美元，月之暗面发布 Kimi K3.1 Pro，中国出台 AI 就业政策

China AI Weekly Digest summarizes the week’s big news, model updates, compute, policy, products, and research, drawn from the China AI Index newsfeed. The digest is (mostly) compiled and drafted by an AI agent using credible news outlets; I fact-check and edit it. Read to the end for “One to Watch” as this week’s video pick.

This week: 306 stories tracked (0 editor picks, 90 English-language, 216 Chinese-language) over 2026-10-16–2026-10-23.

The Big Four

DeepSeek reportedly raises roughly $15 billion, with Tencent and CATL among its biggest backers. Reuters reports that DeepSeek’s latest round has reached about RMB100 billion, against a RMB500 billion pre-money valuation. Tencent and CATL are among the largest investors; Meituan, Geely and GigaDevice are also participating, according to its sources. Reuters says completion is expected in the coming days and the terms could still change. (Reuters/Bloomberg)

China moves to adapt employment policy to AI development. The State Council’s General Office issued guidance setting out 18 tasks across six areas, pairing AI-driven job creation with measures to manage displacement. The priorities include developing new occupations, helping smaller businesses retain and expand their workforce, and assessing the employment risks of AI adoption. The guidance also calls for AI training, support for workers changing roles, and stronger labor protections and social insurance for flexible and gig workers. Alongside these measures, it proposes using AI to improve employment services and establishing monitoring and early-warning systems for job-market disruption. Together, the measures signal that Beijing wants employment considerations built into AI adoption, with support for both new opportunities and workers whose jobs are changing. (财联社)

Moonshot AI launched Kimi K3.1 Pro with a one-million-token context window, alongside a Highspeed version it says runs two to three times faster. The update targets practical work like building software, creating editable 3D assets and conducting financial research. In Moonshot’s reported Terminal-Bench 4.0 results, K3.1 Pro scored 65.7%, ahead of GPT-6 Astra’s 58.2% and just below Claude Opus 5.5’s 66.4%. The model reports 55 on artificial analysis intelligence index, the fourth best model. The company also proactively curbs hacking ability in K3.1 by removing cyber-related data in mid- and post-training. Both versions are available through Kimi’s products and API, with full model weights scheduled for release by November 4. (Kimi technical blog)

Washington restricts DeepSeek in civilian government systems while Beijing keeps a route open for model exports. A Department of Home Security order made public this week bars covered DeepSeek products from federal civilian procurement and requires their removal from agency systems, including certain third-party implementations. Meanwhile, China’s proposed AI export controls remain under consultation through November 15, with provisions facilitating open-source releases and excluding technology already in the public domain. State media reports that DeepSeek and MiniMax have completed registration for open-sourcing and overseas licensing. (AP/SCMP/Recode China AI)

Models

Chinese AI labs are churing out flash-model updates and inference tools. Plus, K3.1 is coming!

DeepSeek released DeepInfer on October 20, making its internally developed inference framework available to other developers. The framework supports V4.1-Flash on both Nvidia’s SM100 Blackwell GPUs and Huawei’s Ascend 950 chips. The company says its production system serves 155 trillion tokens daily from a single cluster, generating $38 in revenue per GPU-hour even at off-peak pricing. (SCMP/@deepseek_ai)
MiniMax launched M3.1 Flash and acknowledged the anonymous “Space Bunny” as an earlier version. Its launch announcement lists API prices of $0.15 per million input tokens and $0.60 per million output tokens. VentureBeat reports strong business-task results alongside a high hallucination rate. (VentureBeat/@MiniMax_AI)
Ant announced open weights for Ling-3.0-flash-Sante, its healthcare-focused mixture-of-experts model, reporting 124 billion total parameters and 5.1 billion active parameters per token. It also announced five Ming-3.0 speech models spanning synthesis, audio understanding and speech recognition, and previewed Ling-3.1-flash-1019, claiming improvements on Artificial Analysis intelligence and agent benchmarks. (@AntLingAGI/@AntLingAGI/@AntLingAGI)
Tencent announced Hy Memory 2.0, including a memory model, framework and MemHub, a CLI and Mac tool intended to synchronize memories across AI applications. (@TencentHunyuan)

Compute

Chip orders, production constraints and semiconductor financing showed how much infrastructure still sits behind cheaper AI.

ByteDance, Alibaba and Tencent reportedly ordered more than three million Nvidia RTX PRO 5500 cards, with a combined sticker value above $35 billion, Reuters reports. These are orders for export-compliant workstation GPUs, primarily for inference—not confirmed deliveries. Nvidia reportedly cannot guarantee fulfillment because demand exceeds planned production. (Reuters)
Cambricon’s third-quarter profit rose 148% year on year to about $210 million, SCMP reports. Yet Bloomberg highlighted sales below expectations and output constraints, adding a production-side qualification to the earnings growth. (Bloomberg/SCMP)
CXMT’s profit turnaround, supported by rising memory prices, did not prevent its shares from falling more than 6%, according to CNBC. (Bloomberg/CNBC)
Alibaba’s DAMO Academy unveiled an automotive RISC-V roadmap and the Xuantie E908A processor IP. Its next-generation C950A targets autonomous-driving and embodied-AI workloads; the roadmap’s performance targets should be distinguished from shipping products. (量子位)
HuiXi Intelligent reported more than RMB2.5 billion in cumulative financing this year as it expands its embodied-AI chip business. That figure is a year-to-date total, rather than the size of a single newly disclosed round. (36Kr)
China’s draft technology-export catalogue also covers AI and memory-chip design and computing systems exceeding 10,000 accelerator cards. The proposal includes facilitation for specified activities, including overseas fabrication and Chinese companies’ own overseas computing centers; the final rules remain unsettled. (Recode China AI)

Funding

Outside DeepSeek’s proposed round, personal assistants and robotics components attracted disclosed capital.

Today AI disclosed more than $200 million in cumulative financing for its personal-AI business. Its product direction centers on persistent memory and assistance that improves with a user’s history. (36Kr)
YITA Dynamics completed seed and seed-plus financings totaling more than RMB100 million. Investors include Matrix Partners China, SenseTime Guoxiang Capital, CCB International, Galbot and Fourier. The company develops embodied-intelligence operating platforms and power systems, putting this investment in the robotics component stack. (投资界)

Policy

Alongside cross-border restrictions, Beijing added institutional support for data development and AI-related employment.

China launched National Data Group in Xiongan, a state-owned enterprise tasked with developing and utilizing data resources and supporting AI development. Vice-Premier Ding Xuexiang attended its inauguration, placing a new corporate institution behind the national data-industry agenda. (Bloomberg/SCMP)
The State Council’s General Office issued guidance on adapting employment policy to AI development, including expanding employment opportunities in emerging fields. It brings labor-market adjustment into the policy discussion alongside the push to deploy AI across industries. (财联社)
The US sanctioned five Chinese entities, including MiraclePlus and Shanghai Zhisuan Technology, in connection with suspected trade-secret theft. SCMP notes that former Google engineer Ding Linwei’s theft convictions remain, while a judge acquitted him on separate economic-espionage counts. (SCMP)

Products

Agents gained new ways to build and search, while access permissions continued to constrain what they could do.

Qwen Go added “My Apps,” allowing users to describe, build and share applications with persistent data. Alibaba Cloud’s Meoo handles application creation, database configuration and deployment. Both the feature and Qwen Go’s cloud Pro capabilities remain in invitation testing. (雷峰网)
Volcano Engine began targeted enterprise trials of Seedance 2.5 4K with 20 studios and directors. The workflow supports generating 480p drafts before regenerating selected shots in 4K. This is a limited professional-production trial, rather than general availability of the 4K model. (雷峰网)
Unitree unveiled H2-X through a promotional film showing protection and rescue scenarios, including lifting a car and confronting an attacker. The demonstrations illustrate its proposed “guardian” role; they do not establish autonomous safety performance. SCMP reports that release details remain sketchy. (SCMP)
Doubao’s phone assistant expanded its refusal-to-operate list from 155 to 166 applications. The update highlights a practical limit on cross-app agents: an assistant’s ability to navigate a screen does not settle whether another platform permits that interaction. (虎嗅)
Pony.ai, inDrive and Nvidia signed an MoU to explore robotaxi testing in Kazakhstan in 2027. The agreement describes planned trials rather than an operating commercial service. (TechCrunch)
Tencent upgraded Yuanbao Search and opened it to developers and enterprises through Tencent Cloud. The company describes changes to compression, indexing and content access, turning an internal search capability into an externally available service for other applications. (36Kr)

Research

Robot learning and verifiable agent environments featured prominently in the week’s research coverage.

Tsinghua, AgiBot and Infinigence introduced Collective Experience Learning, which uses shared robot experience to improve a common policy. In a reported 96-robot experiment, phone-packaging success rose from 79.2% to 98.3%. Humans still labeled outcomes and reset environments; the method removes additional expert action corrections, not all human involvement. (36氪)
Manifold AI’s WorldScape R²SI reached a reported 46.58 on RoboDojo’s Agent Track. Its approach combines task execution with review and revision of reusable skills, with model training also used for some tasks. The result concerns a defined benchmark, rather than general-purpose robot autonomy. (雷峰网)
Tencent introduced WorkForge, a framework for constructing verifiable training environments for agents performing extended work tasks. It starts from expert workflows and builds workspaces and checkable tasks. (@TencentHunyuan)

One to Watch

Subscribe now