The Number
- DeepSeek's off-peak cache-hit input runs $0.003/M tokens on deepseek-flash (DeepSeek-V4.1-Flash), against a $0.15 cache-miss off-peak rate, per its pricing docs (2026-09-23).
- Qwen3-max lists at 2.5 yuan/M input for requests up to 32K tokens, its cheapest max tier, per DashScope pricing (2026-09-23).
- z-ai/glm-5.2 and qwen3.8-27b both sit at $0/M on OpenRouter's free tier, per OpenRouter (2026-09-23).
Top of the Fold
- Alibaba unveils Qwen Intelligence at Yunqi, positioning Qwen as the on-device Agent base for Honor's Magic9 as its first shipping partner, per 智东西.
- DeepSeek publishes a Liang Wenfeng-bylined paper detailing its DSec Agent-training sandbox, disclosing production units of ~160 CPU nodes and 30,000 cores, per 智东西 (preprint, submitted 2026-09-19).
- Xiaomi's model, led by 罗福莉, takes the top spot on a global open-source ranking through large-scale RL, per 新智元, a company-adjacent claim to verify at the leaderboard.
Untranslated
- Alibaba's Qwen Intelligence ships three phone-side agents at launch, Mobile Planner, Mobile-Use, and Mobile Creative, covering task planning, cross-app execution and imaging, per 智东西.
- 夕小瑶 tests MiMo-V2.6 and calls the aesthetic output a real step up for domestic models, one named enthusiast read, hard claims unverified, per 夕小瑶科技说.
- Physical-AI startup 息壤开物 (XIRRA), founded by two ex-Huawei model leads, closes hundreds of millions of yuan across seed and angel rounds in two months to build a native Large Physics Model, per 智东西.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- 息壤开物's founder Li Yin was Huawei Cloud's large-model CTO and co-founder Zhang Hanwang is an NTU professor in causal machine learning, a pairing betting on a physics-first base model rather than an LLM, per 智东西.
The Kicker
- Douyin's 豆包工作 agent adds a Goal mode that keeps checking its own output against acceptance criteria before delivering, a machine that refuses to hand in sloppy homework, per 雷峰网.
