The Number
- DeepSeek bills cache-hit input at $0.003 per million tokens off-peak, a fiftieth of its off-peak cache-miss rate, as of 2026-09-26, per its pricing docs.
- Alibaba's DashScope lists qwen3.8-max-prime at RMB 24 per million input tokens on the on-demand rate, the priciest Qwen text tier it publishes, as of 2026-09-26, per its pricing page.
- Qwen's qwen3.8-27b:free posts $0 per million prompt tokens on OpenRouter, the lowest China-lab listing on the platform, as of 2026-09-26.
Top of the Fold
- Google TPUs ran Kimi 57% faster than Nvidia GPUs in a vLLM-alumni startup's self-reported benchmark, with a DeepSeek inference framework doing the serving, per 量子位.
- 智元 hands over its 20,000th embodied robot and puts 300-plus of them to work across seven venues at Chimelong's new Hengqin park, per 雷峰网.
- Otherwise a quiet morning up top.
Untranslated
- 索辰科技 and its strategic investee 美梦空间 jointly ship an embodied model and a physical evaluation standard, a Chinese-only pairing with no English write-up yet, per 量子位.
- DeepSeek puts a desktop app into circulation with no launch announcement, per 极客公园.
- An FSD-grade team drops its first Physical AI model, Simate-beta, wiring training, inference and evaluation into in-house infra and landing on the RoboDojo board, per 量子位.
The Backstory
- miHoYo's 刘伟, known as 大伟哥, told the 云栖大会 crowd to come slap his face in a year or two if the game studio's AI bet does not land, per 量子位.
The Kicker
- A laptop runs a 700-billion-parameter GLM by treating its SSD as video memory, putting the open project Colibrì at the top of GitHub's trending list, per 量子位.
