The Number
- Baidu lists ernie-5.5-flash-preview at $0/MTok prompt and completion, the cheapest China-lab tier on OpenRouter as of Nov 22, per OpenRouter.
- Qwen3.7-flash runs $0.03/MTok prompt on OpenRouter as of Nov 22, undercutting every DeepSeek flash variant listed, per OpenRouter.
- Tencent's Q3 capex hit 814 billion yuan, up 527% year on year and already above its full-2025 spend, per its earnings filing.
Top of the Fold
- Tencent's Q3 Non-IFRS operating profit fell 6% to 68.1 billion yuan as AI capex and R&D ate the margin, marking a heavier phase of its AI build, per 极客公园.
- Tencent AI chief Yao Shunyu ran the Hunyuan team's first global AMA on Reddit LocalLLaMA, answering 300-plus questions on architecture and leaderboard gaming, days after Hy4 flash shipped at one-twentieth the per-task cost of Hy4, per 极客公园.
- DeepSeek closed a second roughly-100-billion-yuan round with 31 backers buying into Liang Wenfeng's holding vehicle, per 雷峰网.
Untranslated
- A Chinese survey of 80 scholars maps multimodal-agent memory into 297 architectures but only 17 self-evolving ones, framing memory without evolution as the field's gap, per PaperWeekly (preprint).
- A circulating paper argues large models can rewrite their own weights without training and grow stronger, pitched as emergent self-modification, circulating and unconfirmed, per PaperWeekly.
- Dawa Technology, a Chongqing film-set operator turned AI-video lab, raised roughly 500 million yuan across Pre-A and A rounds this year on a control-first Latent Stage base model, per 极客公园.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- Dawa Technology ran physical film sets from 2014 before pivoting to AI in 2023, betting on a "heavy control, light rendering" stack that splits character, motion and lighting control into its own layer, per 极客公园.
The Kicker
- Musk received a figurine made from scrap offcuts, handcrafted by workers at Tesla's Shanghai plant, per 极客公园.