The Overnight

The Overnight — 2026-11-05

Tencent ships Hy4, Baidu drops a driving world model, and foreign labs build on MiniMax

What China's AI world did while the West was asleep. · 2026-11-05

Share: Share on X LinkedIn

The Number

  • DeepSeek's deepseek-flash input runs $0.003/MTok off-peak on a cache hit, half the peak rate, per DeepSeek's pricing docs 2026-11-04.
  • z-ai/glm-5.2 lists at $0.006/MTok prompt on OpenRouter, the cheapest China-lab prompt tier there 2026-11-04.
  • Qwen3.8-Max runs 12 yuan per 1M input tokens on Alibaba's Bailian, with Batch calls at half that, per DashScope pricing 2026-11-04.
  • Kimi K2 series stops onboarding new users from Oct 31, 2026, with full retirement planned by late December, per Moonshot's pricing docs 2026-11-04.

Top of the Fold

  • Tencent ships Hy4 in final form with vision added for the first time and plans to open the weights, pitching it against DeepSeek-V4.1-Flash on agentic coding and search, per 智东西.
  • Baidu Apollo releases Apollo World, an agent-driven full-modal world model it says lets robotaxis clock 100,000 virtual km a day in the cloud, a company-stated figure, per 新智元.
  • Nvidia's technical team and Pika both build on MiniMax's open H3 video model, with licensing attached despite public weights, a sign Chinese open releases are becoming foreign R&D starting points, per 智东西.

Untranslated

  • 招商局狮子山实验室 releases 驭物 Lion-ER-35B-A3B, an embodied-reasoning model activating only 3B parameters that it claims improves across all ten benchmarks, a company-stated result, per 量子位.
  • 虎鲸文娱 launches 鲸锐AI, an AI film-production platform wiring in 50-plus third-party video models and 10-plus in-house ones, claiming over 30% cost savings per finished minute, per 雷锋网.
  • Ant's 灵光 expands to 40 external connectors overnight, adding 微信读书, 夸克 drive, 哈啰 ride-hailing and McDonald's, pushing its agent from digital tools into daily errands, per 雷锋网.
  • 复旦, 香港城市大学 and 香港中文大学 release LoLBench, dropping agents into codebases averaging 2.4M lines where even top models deliver only 14% of a 5,000-word spec, per preprint arXiv:2609.37143.

The Backstory

  • 尹希, once among Harvard's youngest Chinese full professors and a CUSTC 少年班 alumnus, leaves OpenAI after just five months, writing he is ending the journey early to explore a new direction, per 智东西.

The Kicker

  • DeepSeek's Harness quietly pushed a v0.2.2-alpha.3 pre-release the same day its web and API partly went down, keeping its community's half-joking rule intact: every crash comes with an update, per 智东西.