The Number
- DeepSeek-V4.1-Flash lists at $0.15/M cache-miss input off-peak, roughly half its cache-miss peak rate, per DeepSeek's pricing docs as of 2026-09-12.
- Qwen3-Max runs 2.5 yuan/M input under 32K tokens, the entry tier on DashScope, per Alibaba's pricing page as of 2026-09-12.
- GLM-5.3-Flash sits at $0.075/M prompt on OpenRouter, mid-pack among China-lab flash tiers as of 2026-09-12.
Top of the Fold
- Moonshot drops Kimi K2.8 with million-token context opened to all users and performance it says nears K3, landing as the company sprints toward a Hong Kong IPO, per 量子位 (company-stated benchmarks, unverified).
- Zhipu launches a fresh $5B raise, a $2B share placement plus $3B convertible notes, just two months after its last round, signaling the capital war is intensifying even as fewer labs remain, per 虎嗅 citing Reuters.
- 生数 shows Motus2, a self-evolving world model that lets robots predict and grade the outcome of a dexterous action before refining strategy, demoed live at the Bund Summit, per 新智元 (company-stated).
Untranslated
- A leaked WeChat group screenshot with no source spread across China's VC and AI circles overnight, forcing 月之暗面 to post a legal statement calling the claims about its founder and staff fabricated and saying it has filed a police report, per 凤凰网财经 via 虎嗅 (circulating/unconfirmed).
- 蚂蚁 rebrands its GPASS device-trust framework to 灵影, positioning it as an agent-native OS for AI glasses and new terminals and opening it to chip and hardware partners, per 雷峰网.
- Chinese AI product names are quietly sinicizing, with Alibaba's QwenWork becoming 千问办公 and Baidu's Dumate leaning on a Chinese 搭子 brand, one 虎嗅 columnist's read that the geek-era English names no longer travel to mass users.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- At the Bund Summit, 22-year-old 陈博远 shared a stage with veteran founders, having built 逆矩阵科技 during his Peking University undergrad and raised a seed-plus round north of $100M in half a year, per 雷峰网.
The Kicker
- The Bund Summit's AI coding hackathon drew over ten thousand entrants, nearly 70% of them minors, with the youngest competitor aged six, per 雷峰网.
