NaiveAI, founded by Tsinghua's Dai Jifeng, raises $400M in 7 months betting on post-training

创业邦 · wechat · 2026-09-23

NaiveAI, founded 7 months ago by Tsinghua associate professor Dai Jifeng (70k+ citations, creator of InternVL), has raised $400M across three rounds at a $1.42B valuation from Tencent, Sequoia China, IDG and Matrix Partners.

The bet: skip pre-training. Instead of the capital-heavy from-scratch route taken by DeepSeek and Moonshot, NaiveAI fine-tunes an existing Chinese open-source model — architecture tweaks, mid-training, then RL — cutting iteration costs by an order of magnitude. CMU research supports the undervalued role of mid-training. Long-term target: recursive self-improvement.

Context: the path has precedents — Thinking Machines Lab built on DeepSeek-V3's architecture, and Cursor's Composer-1 shows traces of Zhipu's model. But commercialization pressure is brutal: all Chinese AI models combined earn 10% of OpenAI's annualized revenue. Zhipu made $954M RMB in H1 (86.5% from API), MiniMax $120M. Mistral shows open-source monetization can work ($20M to $400M ARR via private deployment contracts).

Risks: an unresolved IP dispute with MiroMind, and the question of how much differentiation post-training can deliver over someone else's base model.

Original post →

More from Companies & People

Companies & People channel →