The Number
- DeepSeek-V4.1-Flash-1017 (Max) sits as the highest-ranked open model on Code Arena's WebDev board at 1670, 12 behind GPT-6 Sol (Max), on Arena's preference ranking as of 2026-10-19, per Arena.
- qwen3.7-flash lists at $0.03/M input, the cheapest paid China-lab model on OpenRouter, as of 2026-10-19, on OpenRouter.
- Zhipu's glm-5.3-flash lists at $0.048/M input on OpenRouter, as of 2026-10-19, on OpenRouter.
Top of the Fold
- MiniMax ships M3.1 Flash, a 1M-context coding and agent model priced at ¥1/M input and ¥4/M output, pitched as matching DeepSeek's V4.1 Flash on agentic tasks (company-stated), per 智东西.
- MemSilicon closes a ¥400M-plus A round and names ex-Graphcore China chief 卢涛 CEO, pivoting from RRAM storage into a first-generation LPU inference chip, per 雷峰网.
- DeepSeek opens standalone Search and Fetch APIs at $0.75 per 1,000 requests, extending its stack from models into web retrieval, per DeepSeek docs.
- Ant leads a C round in Tsinghua-rooted AI-fintech 熵简科技, tying the firm into Ant's financial-agent and vertical-model work, per 智东西.
Untranslated
- Headhunters describe an AI pay inversion across BAT, with junior base-model staff out-earning senior colleagues and talk of scrapping job levels, circulating and unconfirmed, per 雷峰网.
- AliExpress says it will take angel stakes in AI-hardware startups, aiming to incubate 100-plus brands before their products even ship, per 雷峰网.
- Lovart ships 2.0 with Arti, a personal design agent built to remember a user's taste across projects, per 智东西.
The Backstory
- At Shanghai Jiao Tong, 贾萧松's PhD swung from autonomous-driving trajectory prediction to end-to-end driving after a dorm-room phone call from advisor 严骏驰, who argued prediction was boxed in between perception and planning, per 雷峰网.
The Kicker
- Tsinghua's embodied-AI final now grades students on whether a robot they trained can fold a shirt on the spot, part of a 1,000-robot campus rollout, per 新智元.
