DeepSeek 发布 V4.1-Flash，燧原科技上市，Anthropic 指控中国实验室蒸馏 Claude

China AI Weekly Digest summarizes the week’s big news, model updates, policy, products, and research, sourced from the aggregated newsfeed of China AI Index. The digest below is mostly compiled and drafted by an AI agent sourcing from credible news outlets, then fact-checked and edited by me.

This week: 297 stories tracked (4 editor picks, 89 English-language, 204 Chinese-language 🇨🇳) over 2026-09-05–2026-09-12.

The Big Three

DeepSeek had one of its biggest weeks yet: a new flagship model and a $74 billion pre-IPO valuation. The Hangzhou lab shipped DeepSeek-V4.1-Flash, a 552B-parameter MoE model that activates as few as 8–16B parameters per token and, on Terminal-Bench 2.1, edged out both OpenAI’s GPT-5.6 Sol and Moonshot’s Kimi K3 on coding and agentic tasks. In the same week, DeepSeek reportedly hired Citic Securities and three other underwriters to prepare a Shanghai STAR Market listing, with a pre-IPO round reportedly valuing the company near ¥500 billion (~$74 billion). (Reuters/SCMP/FT)

China’s AI chip IPO wave crested with Enflame’s blockbuster Shanghai debut. The Tencent-backed Nvidia challenger raised ¥6.12 billion in its STAR Market IPO, then saw shares peak at a 234% gain before closing up 188%, valuing the company near $25.5 billion. Enflame is the last of China’s “four AI chip dragons” (alongside Moore Threads, Biren, and MetaX) to go public, and the frenzied reception underscores how much investor appetite Beijing’s semiconductor self-sufficiency push has generated — even as the debut immediately reignited debate over whether China’s humanoid-robot and chip IPOs are running ahead of fundamentals. (Bloomberg/Reuters/SCMP)

Anthropic accuses Chinese AI labs of distilling Claude to train their own models. Anthropic said it disrupted campaigns in which the three Chinese labs routed millions of user queries through Claude to harvest training data. Washington quickly amplified, with US officials calling it “industrial-scale” theft of AI technology just two weeks before Trump and Xi are set to meet. Beijing rejected the allegations outright and warned of retaliation. (Bloomberg/Reuters/SCMP)

Models

DeepSeek’s Flash launch dominated the model news, but agentic tooling and a couple of open-source drops from elsewhere in the ecosystem were the other thread this week.

• DeepSeek shipped V4.1-Flash, the smallest model in its new architecture family with native visual understanding, claiming benchmark wins over GPT-5.6 Sol and Kimi K3 on coding and cybersecurity tasks. (VentureBeat/SCMP)

• ByteDance, Alibaba, and Tencent are all pivoting spending toward “AI digital employee” office agents. Tencent’s WorkBuddy opened its agent platform to third-party developers, while Doubao pushed its own office-agent product. (虎嗅/第一财经)

• Alibaba’s Qwen upgraded its Token Plan for individual developers with 12 new Agent Harness tools, and Perplexity reportedly built a local agent product on top of Qwen3.8 rather than a US model. (雷峰网/雷峰网)

• Infinigence and Genesis Motion jointly released NeoHorse, described as the first “Agent-Native” model exploring Harness-driven recursive self-improvement. (极客公园)

• Tencent Hunyuan open-sourced AuK, a foundation model for unified zero-shot speech generation and editing. (@TencentHunyuan)

• Ant Group open-sourced Ling-3.0-flash-VL, a vision-language model shipping in BF16 and FP8 with FP4/INT4 planned. (@AntLingAGI)

Funding

The week’s real story was IPOs.

• Moonshot AI (Kimi) is targeting roughly $2 billion in annualized revenue in 2026 as it heads toward a listing. (TechCrunch/Bloomberg)

• Moonshot is now exploring a dual Hong Kong–Shanghai listing rather than a single venue, sources told SCMP. (Reuters/SCMP)

• French investors reportedly bought into Moonshot at a $30 billion valuation ahead of the Kimi IPO. (36氪)

• Z.ai (Zhipu) filed for a $5 billion Hong Kong share sale plus convertible bonds, per term sheet details seen by Reuters. (Reuters)

• Alibaba is backing an AI-testing startup founded by a former employee at a $2.5 billion valuation. (Bloomberg)

• China’s AI chipmakers are raising prices as a high-bandwidth-memory shortage bites into supply, a headwind for every company riding the domestic-chip IPO wave above. (Reuters)

• Following Unitree’s volatile trading debut, regulators are reportedly curbing further humanoid-robotics IPOs while the sector’s valuations — split roughly ¥50–150 billion — get recalibrated. (The Information/虎嗅)

Policy

The distillation dispute (see Big Three) doubled as this week’s policy story, but a few more developments landed alongside it.

• South Korea is toughening its spy law specifically to protect chip secrets from China, as regional governments harden IP defenses. (Bloomberg)

• Nikkei reports the US will press China on AI-directed cyberattacks at the September 24 Trump-Xi summit, with AI guardrails also on the agenda. (Nikkei Asia)

• China’s Supreme People’s Court issued new AI “red lines” covering deepfakes and privacy protections. (SCMP)

• Beijing set a target to quadruple national AI computing capacity by 2030 as part of a broader tech push. (SCMP)

• Some Chinese AI firms reportedly avoided meeting a visiting US delegation this week over fears of drawing fresh sanctions scrutiny. (SCMP)

Products

• Alibaba’s AI “employees” can now operate inside third-party apps, including ByteDance’s and Tencent’s, a notable cross-platform interoperability move between rivals. (SCMP)

• Moonshot’s Kimi and Zhipu’s Z.ai both launched AI subscriptions for sale directly on Tmall, moving the model-subscription race onto retail shelves. (SCMP)

• Xiaomi unveiled a folding phone built on a home-grown chip, positioning it directly against Apple and Huawei. (SCMP)

• JD.com is committing to deploy 3 million robots to automate its logistics network as part of the broader AI shift among Chinese platform companies. (SCMP)

• Unitree fully open-sourced its UnifoLM-WLA-1.0 embodied foundation model and demonstrated a world-model-driven humanoid robot performing fully autonomous combat/boxing moves — the latter corroborated by mainstream Chinese outlets. (第一财经/@UnitreeRobotics)

• WeRide received Spain’s first national L4 robotaxi operating permit alongside Uber. (雷峰网)

• WeRide says it also began fully driverless L4 robobus operations at Zurich Airport with no safety operator onboard. (@WeRide_ai)

• Pony.ai began Europe’s first fully driverless robotaxi test rides, in Zagreb, with no operator behind the wheel. (@PonyAI_tech)

Research

Lighter on formal research this week, but two items worth flagging.

• SCMP argues China should study Ukraine’s wartime PR and information-operations playbook as AI reshapes strategic communications. (SCMP)

• ByteDance’s founder is reportedly personally racing to build a competitive “world model,” joining the small group of AI leaders chasing the technology. (Bloomberg)

Subscribe now