The Number
- DeepSeek-V4-Pro cache-miss input runs $0.435/M off-peak versus $0.87/M output, before a peak/off-peak split takes effect 16:00 UTC on Aug 16, per DeepSeek's pricing docs.
- Qwen3.8-max sits #8 on Arena's text preference ranking at Elo 1491, the top China-lab entry, as of Aug 15, per LMArena.
- Kimi-k3-max holds #2 on Arena's coding preference ranking at Elo 1674, behind only Claude, as of Aug 15, per LMArena.
- Qwen3.7-flash lists at $0.03/M prompt on OpenRouter, the cheapest China-lab model there, as of Aug 16.
Top of the Fold
- DeepSeek moves to peak/off-peak API billing from 16:00 UTC Aug 16, halving off-peak rates so V4-Flash input drops to $0.22/M cache-miss, the first time a Chinese lab prices by time of day, per DeepSeek's docs.
- Zhipu ships GLM-5.3 on the same GLM-5.2 base, claiming a 50% internal coding gain and open-source first on TerminalBench 3.0, with weights promised in two weeks, per its release (company-stated, unverified).
- Apple has trained a China-market LLM with Alibaba, a shift from relying on third-party models, per 极客公园 citing three people familiar; circulating, unconfirmed.
Untranslated
- 至知研究院 pitches an interpretability route that reads model behavior by decomposing weights rather than training a surrogate network, claiming data cost under 1%, per 量子位; company-stated, verify.
- A DeepSeek Harness plugin trended on GitHub overnight, users bolting on long-term memory, a virtual pet, and mini-games onto the model, per 量子位.
- 音潮 opened its music-generation API free for a limited run, positioning directly against Suno on the flaw Chinese users complain about most, per 量子位; company-stated.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- Tencent's new AI products (Hunyuan, Yuanbao, CodeBuddy, WorkBuddy) cut about 10.5B yuan from Non-IFRS operating profit this quarter, up from roughly 8.8B in Q1, meaning the legacy business grew 19% while the AI bet dragged the reported figure lower, per Tencent's Q2 filing as read by 极客公园.
The Kicker
- A rough-cut animated short called 《牛来》 is spreading on Douyin precisely because it looks unpolished, viewers reading the crude models and stiff motion as proof a human actually finished it, per 虎嗅.
