The Number
- DeepSeek-V4-Flash lists at $0.14/M cache-miss input and $0.28/M output as of Aug 7, with the docs warning of a coming increase, per its pricing page.
- Kimi-K3-Max sits #2 on Arena's coding preference ranking at Elo 1676 as of Aug 6, behind only Claude Opus-5-Max, per LMArena.
- Qwen3.8-Max holds #5 on Arena's overall text preference ranking at Elo 1496 as of Aug 6, the top-rated Chinese entry, per LMArena.
- Qwen3.7-Flash routes at $0.03/M input on OpenRouter as of Aug 7, the cheapest China-lab prompt tier on the platform.
Top of the Fold
- DeepSeek tells API users it plans to raise pricing significantly in the near term, breaking from the race-to-zero as Meta and OpenAI wade into cheap inference, per its pricing docs.
- Meta ships its Muse Code agent and Muse Spark 1.2 priced below DeepSeek-V4-Flash, with a further 95% cut for users who let it train on their data, landing the same day DeepSeek signaled its own hike, per 智东西.
- DeepSeek is said to have reopened a second funding round targeting 50B yuan at a ~500B-yuan pre-money, circulating and unconfirmed, aiming to sign in late August, per 财经 via 极客公园.
Untranslated
- USTC built a robot catalysis lab of 45 modular workstations and had six agent frameworks run end-to-end experiments in it, pushing the autonomous-science test from Q&A into real physical execution, per 新智元 (preprint, arXiv:2607.23045).
- A Chinese team converted 21 of 28 attention layers in Qwen3-0.6B-Base to Kimi Delta linear attention on a single consumer GPU, then documents exactly what the surgery breaks, per arXiv:2608.02689 (preprint).
- Otherwise quiet across the Chinese community overnight.
The Backstory
- Insta360 founder Liu Jingkang reframed the camera business as a "Camera Agent," already selling cloud AI editing at 6 yuan a clip, arguing taste matters more than task execution, as recorded at AGI Playground 2026, per 极客公园.
The Kicker
- Honor previewed Android's first "liquid glass" UI for MagicOS 11, opening beta signups Aug 6, per 智东西.
