The Number
- DeepSeek V4-Pro now runs $0.66/M input at off-peak cache-miss, with cache-hit input at $0.003/M, per DeepSeek's docs as of 2026-10-12.
- GLM-5.3-flash sits at $0.0232/M prompt on OpenRouter, the cheapest China-lab flash tier listed, per OpenRouter as of 2026-10-12.
- Qwen3.8-max lists at 12 yuan/M input with a 1M-token free allowance in Beijing only, per Alibaba DashScope pricing as of 2026-10-12.
Top of the Fold
- Infinigence (无问芯穹) closes a 2.3B-yuan C round, the largest domestic AI-Infra raise to date, lifting its total to 4.3B as it preps an IPO, per 智东西.
- Shanghai's MOSI (模思智能) takes over 1B yuan Pre-A led by Huakong, its second raise this year, to fund full-modal real-time interaction, per 智东西.
- Zhipu quietly opens GLM-5.3 HighSpeed to random ZCode subscribers at a tested 490 tokens/s, several times its 70.84 tokens/s standard API, per 智东西.
Untranslated
- Shanghai AI Lab's Harness-of-Harness lets an agent build an FPS game from scratch in 5 days with no human touching code, a self-directed-engineering claim worth verifying, per PaperWeekly (preprint).
- Huawei's RSI agent clears 25 games including a 21:0 run, using 56% fewer actions than humans and dropping the LLM call entirely late-game, company-stated per 量子位.
- A Qwen3-VL-backed reward model trains with zero human preference labels, scoring web captions above perturbed copies, circulating as a China-lab result on arXiv (preprint).
- Over 40,000 ICLR submissions are being traded like stocks in a Chinese researcher game, buying papers long or short on acceptance odds, circulating per PaperWeekly.
The Backstory
- MOSI traces to Fudan professor 邱锡鹏's OpenMOSS team, which shipped China's first ChatGPT-style open model MOSS in February 2023, now reborn as a funded startup, per 智东西.
The Kicker
- REA (Reverse Engineer Anything) jumped from under 10,000 GitHub stars on Oct 6 to over 110,000 by Oct 12, mostly because people wanted Halo and GTA5 running in a browser tab, per 极客公园.
