The Number
- Kimi K3 Max holds #2 on Arena's coding preference ranking at Elo 1682, the only Chinese model in the top five behind Anthropic's Claude Opus 5 Max, per LMArena as of 2026-07-30.
- DeepSeek-V4-Pro lists at $0.435/M input on a cache miss and $0.0036 on a cache hit, a two-order gap that rewards heavy context reuse, per DeepSeek's pricing docs as of 2026-07-31.
- Qwen3.7-Flash prices at $0.03/M input on OpenRouter, the cheapest China-lab prompt tier on the platform, per OpenRouter as of 2026-07-31.
- Doubao's daily model-call volume tops 180 trillion tokens, against a stated $4B annualized run-rate for ByteDance's model business, per figures cited by 虎嗅 on 2026-07-30.
Top of the Fold
- Moonshot posts Kimi K3 as its flagship 1M-context model on the official pricing page, moving the K2 series toward retirement and putting a paid frontier tier behind its Arena coding rank, per platform.moonshot.cn as of 2026-07-31.
- ByteDance folds the Feishu product team into Doubao and merges Feishu's go-to-market with Volcano Engine, its largest ToB reshuffle since 2021, betting the enterprise story on one Doubao stack, per 雷峰网.
- Tencent's WorkBuddy ships a human-AI co-writing mode that drops an editable Tencent Docs pane beside the chat, killing the copy-paste loop that has slowed every AI office tool, per 智东西.
- iFlytek debuts a real-time-generated digital human that walks, guides, and switches languages on a transparent screen, pushing digital avatars from marketing prop to store-front interface, per 智东西.
Untranslated
- Moonshot completed a Series F above $3.5B at a $50B valuation, circulating and unconfirmed, which would make it the best-capitalized of the model labs going into K3, per 雷峰网's morning brief.
- A cross-city inference relay using PD-separated KV-cache transfer cuts token cost by nearly 40% with latency roughly hidden, framed as pooling idle national compute, in a technical report flagged by 量子位 and PaperWeekly as of 2026-07-30; treat throughput claims as authors'-stated.
- Evoken founder 陈冕 says the company crossed $3 US ARR across Liblib, Lovart and LibTV yet describes it as one of the most fragile AI-app businesses alive, tomorrow's model release always threatening today's value, as recorded by 晚点聊 on 2026-07-30.
- Delta Intelligence closed a near-500M-yuan angel++ round, its sixth raise in six months, from robot OEMs including 智元 and 乐聚, on a humanoid foundation-model pitch, per 雷峰网.
- A ByteDance-plus-Tsinghua-Yao-class team raised three rounds totaling hundreds of millions of yuan inside a year to rebuild enterprise software engineering, circulating as a fresh funding signal, per 量子位.
The Backstory
- Evoken's 陈冕 tells 晚点聊 he cries when he thinks the company is finished and laughs loudly the moment a new direction clicks, the founder of a $2B-valued AI-app shop describing anxiety as the trait this era selected for, as recorded by 晚点聊 on 2026-07-30.
The Kicker
- Out-of-work drama extras in China are moonlighting as scenic-spot NPCs, one telling 虎嗅 his day rate jumped roughly tenfold to as much as 8,000 yuan for standing around in period costume and posing with tourists.
