The Number
- China labs hold four coding slots on Arena's preference ranking, led by Alibaba's qwen3.8-max-0902 at #4 (Elo 1681), ahead of Moonshot's kimi-k3-max at #5, per LMArena as of 2026-09-21.
- DeepSeek's flash tier runs $0.15/M input on a cache miss off-peak, half that on cache hit, per its pricing docs as of 2026-09-22.
- Four Chinese models each cleared 10 trillion tokens in a single week, all lightweight Flash-class, per 智能经济30人论坛's AGI Market Watch (Sept 12-18) as reported by 雷峰网 on 2026-09-22.
Top of the Fold
- Alibaba Cloud lays out an Agentic Cloud stack at Yunqi with AgentCore, Agent Sandbox and a new CPFS storage tier, betting enterprise agents need compute and governance more than another model, per 雷峰网.
- Xiaomi's MiMo team open-sources MiMo-V2.6, two native omni-modal models plus the RL training framework, framing it as the largest single RL run of any open model to date, per 智东西.
- Tencent Hunyuan ships Hy Image3.5 preview and puts it inside Tencent Video's WorkRally platform free for two weeks, aiming the model straight at film storyboarding and shot production, per 雷峰网.
- Alibaba says Qwen4 and its next video model are both in training, with an Alibaba researcher stating post-Qwen4.5 versions will scale to 5-10T parameters, per 量子位.
Untranslated
- DeepSeek's Sept 10 price cut on V4.1 Flash pushed its weekly token use up 483.62% and its ranking from #9 to #2, with Tencent's Hy4 preview third at 12.01T tokens, per 智能经济30人论坛's AGI Market Watch as reported by 雷峰网.
- Xiaomi's Luo Fuli says MiMo-V2.6 spent six days and over ¥20M on a public 'live training' run, and calls the effort a step toward recursive self-improvement, per 智东西; the RSI framing is her stated view, not a demonstrated result.
- 浪潮信息 says it ran Moonshot's 2.8-trillion-parameter Kimi K3 on a single machine at 5.85ms first-token, claiming 10x token throughput per unit of investment, per 智东西; company-stated, and K3's 1.56TB weights normally need 64-plus cards.
- SenseTime releases U1 Pro image generation, with 量子位 noting its outputs were hard to spot as AI-made, let alone domestic; treat the quality read as circulating.
The Backstory
- 华为 unveils what it calls the world's first 3D data center, stacking cooling, IT, power and backup as separate vertical layers borrowed from chip-fab design to sell AIDC as a product rather than a custom build, per 智东西.
The Kicker
- Baidu Tieba users spent the night of Sept 16 racing a midnight deadline to upload hundreds of terabytes from the shuttered '失传媒体吧' forum to the Internet Archive, a volunteer rescue of China's lost internet memory, per 虎嗅.
