The Number
- Moonshot's kimi-k3-max sits #2 on Arena's coding preference ranking at 1676 Elo, the only China entry inside the top five as of Aug 2, per LMArena.
- Z.ai's glm-5.2-max holds #6 and DeepSeek's deepseek-v4-flash-high #7 on the same coding board, the next Chinese labs behind Kimi as of Aug 2, per LMArena.
- DeepSeek-V4-Flash lists at $0.14/M input (cache miss) and $0.28/M output as of Aug 3, with a peak/off-peak doubling flagged as coming, per DeepSeek's pricing docs.
- Qwen3.7-Max runs a limited 50% cut to 6 yuan/M input, 18 yuan/M output as of Aug 3, per Alibaba's DashScope pricing page.
Top of the Fold
- Moonshot marks Kimi K3 as officially launched with a 1M-token flagship tier on its platform, its first top-line model since the K2 series, per Moonshot's pricing docs.
- Otherwise a quiet morning up top; the overnight wires ran mostly on US labs.
Untranslated
- ByteDance's Seedance 2.5 and MiniMax's H3 video models both went live on the last workday of July, with Seedance 2.5 API priced ~50% above SD2.0 and capped at 720p for now, per 夕小瑶科技说.
- Seedance 2.5 keeps the unified audio-video generation path and pushes a 30-second single-shot narrative length, though the API model ID is still unpublished and expected about a week out, as described by 夕小瑶科技说 (company-stated).
- Otherwise quiet across the Chinese community overnight.
The Kicker
- A huxiu essay recounts an MIT writing teacher spotting AI-drafted student stories on the first paragraph, calling the too-clean prose "dead perfection," per 虎嗅.
