The Overnight

The Overnight — 2026-08-08

ByteDance's trillion-parameter model, and Kimi K3 climbs Arena code

What China's AI world did while the West was asleep. · 2026-08-08

Share: Share on X LinkedIn

The Number

  • DeepSeek-V4-Flash lists a cache-miss input price of $0.14/M tokens against $0.435 for V4-Pro, with a warning that overall pricing rises soon, per DeepSeek's docs as of 2026-08-08.
  • Alibaba's flagship qwen3.8-max sits at #5 on Arena's text preference ranking at 1497 Elo, the top China-lab entry, as of 2026-08-07.
  • Moonshot's kimi-k3-max holds #2 on Arena's coding preference ranking at 1675 Elo, behind only claude-opus-5-max, as of 2026-08-07.
  • qwen3.7-flash routes at $0.03/M input tokens on OpenRouter, the cheapest China-lab listing there, as of 2026-08-08.

Top of the Fold

  • ByteDance is training a model of up to 10 trillion parameters, still in a 3-to-6-month pretraining run led by Seed's 项亮, a scale that would top Kimi K3 and near Anthropic's Fable 5, per 晚点 LatePost and the FT via 智东西.
  • Alibaba opens public beta of its Wan 3.0 video model, which turns documents and slide decks into video, widening the generation race beyond text-to-clip, per 量子位.
  • ByteDance ships Seedance 2.5, doubling single-shot video length from 15 to 30 seconds, the same day MiniMax open-sources H3, per 新智元.
  • Ant Group open-sources Avernet, a multi-agent collaboration infrastructure, pushing agent orchestration toward production, per 量子位.
  • Alibaba launches CosyVoice Studio, billed as China's first one-stop AI voice platform folding semantic understanding into speech, per 量子位.

Untranslated

  • JD open-sources JoyAI-Video-Edit, applying full-duplex real-time editing to video the way GPT-Live did to voice, per 新智元.
  • LightX2V's LightWan2.2-A14B claims a 14B video model running on a single RTX 5090, with 8-card 720p T2V in 3.8 seconds, a company-stated 705x speedup over the 40-step baseline, per 新智元.
  • 智元机器人 removes chief scientist 罗剑岚 from its site, a quiet executive shuffle circulating unconfirmed, per 量子位.
  • Recursive SuperIntelligence, co-founded by 田渊栋, discusses its spring RSI results at a reported $4.6B valuation on 晚点's LateTalk, framed as small-scale proof a system can auto-improve its own infra, per 晚点聊.

The Backstory

  • 田渊栋 traces the RSI revival to a decade after Google's AutoML neural-architecture search, arguing AutoResearch has only now caught up, as recorded on 晚点聊 episode 178.

The Kicker

  • Chinese parents are watching 10后 kids speak in nothing but internet 烂梗, and the counter-movement is a wave of AI-made clips mocking the slop, per 虎嗅.