The Number
- DeepSeek V4 Pro lists at $0.435/M input (cache miss), $0.87/M output, with docs flagging a coming price rise, per DeepSeek's pricing page 2026-08-13.
- Qwen3.8-max sits at #9 on Arena's text preference ranking (Elo 1490), the top China-lab entry overall, per LMArena as of 2026-08-12.
- Kimi-k3-max holds #2 on Arena's coding preference ranking (Elo 1674), ahead of qwen3.8-max at #3, per LMArena as of 2026-08-12.
- 面壁智能 crossed 200亿元 valuation on 50亿元+ raised in H1 2026 as it files for A-share IPO coaching, per 智东西.
Top of the Fold
- DeepSeek pushes V4 Pro to general availability (version 0813), with its DeepSWE coding-agent score jumping from 12.8 to 62.7 over the preview, a company-stated leap it frames against Claude Fable 5, per 智东西.
- Manus tells users it is leaving Meta to run as an independent company again, closing a multibillion-dollar deal seven months after the December 2025 acquisition, with some post-acquisition user data set for deletion Aug 23-24, per its user letter via 智东西.
- Tencent says it prepaid roughly 514亿元 for compute in Q2, steering new GPU capacity first to its own model training and AI apps rather than rental, per its Q2 earnings call via 智东西.
- Honor opens sales of its Robot Phone at 9,999元 with a 4-DOF arm and a 300B-parameter agent base, claiming 400,000 pre-orders, per 智东西.
Untranslated
- Lin Junyang, ex-Qwen model lead, names his new Shanghai agent lab Pragmatik Labs (语用科技), backed by Hillhouse, Sequoia China, Tencent and a Shanghai state fund at a reported $2B post-money, per 虎嗅.
- Ex-Huawei Noah's Ark lab head Hao Jianye's Shenzhen startup MemoraX AI closed a several-hundred-million-yuan seed++ round, its third raise since March 2026, targeting long-term memory infrastructure for large models, per 智东西.
- Chinese benchmark MLS-Bench, 140 real research tasks across 12 fields, appears in the official launch tables for Kimi K3 and Qwen3.8-Max, and finds frontier models still cannot reliably invent new ML methods even with human solutions in hand, marked preprint, per 新智元.
- Alibaba's Yuvion VL, an 8B multimodal model, is claimed to top a multimodal safety board over a 397B model, a company-stated result to verify, per PaperWeekly.
The Backstory
- Lin Junyang studied English, plus Japanese, Russian, German and French, at the University of International Relations before a foreign-languages master's at Peking University, an unusual path into leading Alibaba's Qwen models, per 虎嗅.
The Kicker
- At Mengniu's internal AI contest, a dairy-unit staffer used Tencent's WorkBuddy agent to narrow a sick cow's likely diagnosis to 1-3 conditions in two rounds of Q&A, per 智东西.
