The Number
- Qwen3.8-Max lists at 12 yuan per 1M input tokens (36 output) in China, 1.5 yuan on implicit cache hit, per Alibaba's DashScope pricing (2026-08-04).
- DeepSeek-V4-Flash holds a $0.14/M cache-miss input price (0.28 output) with a 1M context, per DeepSeek's docs (2026-08-04).
- MiniMax H3 tops Artificial Analysis's audio-video editing board at 1130 Elo, ahead of Gemini Omni Flash and Wan 2.7, per 智东西 (2026-08-03), company-stated.
- Kimi K3-Max sits #2 on Arena's coding preference ranking at 1676 Elo, behind only claude-opus-5-max, per LMArena (2026-08-03).
Top of the Fold
- Alibaba ships Qwen3.8-Max at 2.4 trillion parameters (95B active), its first Max-tier model slated to open its weights, sending Alibaba's Hong Kong stock up 7.26%, per 智东西.
- ByteDance, Alibaba, and Tencent all move on AI office within two weeks, ByteDance folding Feishu's product team into Doubao and Alibaba opening QwenWork to public beta, turning AI from a feature into the base layer of work apps, per 智东西.
- MiniMax open-sources H3, a full-modality video system generating 15-second 2K clips with native stereo audio, adapted by 16 chip and platform partners on day one, per 智东西.
- SenseTime previews SenseNova U1.5-Lite, an 8B open multimodal model that turns hand-drawn sketches into print-standard posters, per 量子位.
Untranslated
- 西湖心辰, the Westlake-affiliated diffusion-language-model lab founded by ALBERT first author 蓝振忠, closes a several-hundred-million-yuan B+ round led by 广发信德 with Ant Group backing, earmarked for a new diffusion language model and overseas push, per 智东西.
- LatePost's tech podcast dissects Kimi K3 with SGLang core dev 赵晨阳, framing it as a 3T open-weight release that opened the weights but kept the pipeline that produced them, per 晚点聊 (attribute; podcast notes, not transcript).
- Huawei's Noah's Ark lab open-sources MindMemOS, a transferable self-evolving memory layer for AI agents so they no longer forget between sessions, per 量子位.
- StepFun chairman 印奇 makes his first public appearance to unveil the STEPX Neo agent phone, and 虎�ff learns the company is running two internal tracks, one model-led and one agent-terminal-led, per 虎嗅 (one outlet's reporting, unconfirmed).
The Backstory
- 西湖心辰 founder 蓝振忠 built the lab out of Westlake University after co-authoring Google's ALBERT and serving as a scientist at Google AI, a research-lab-to-startup path less told in English than the marquee founding stories, per 智东西.
The Kicker
- A Huazhong University team ships AlphaOracle, an AI that mimics an expert's workflow to help read oracle-bone script, cross-referencing bronze inscriptions and Warring States glyphs so a scholar's days-long lookup collapses to one pass, per 新智元 (preprint).
