Tsinghua professor's stealth LLM startup Naive AI hits $1.4B valuation
What's the deal? Naive AIDealroom has a profile for this one. Try Dealroom →, a secretive Beijing-based startup founded in February by Tsinghua University professor Jifeng Dai, raised $120 million in a round that closed in September 2026, reaching a $1.42 billion post-money valuation. Investors include Tencent, IDG Capital, MPCiDealroom has a profile for this one. Try Dealroom →, and HSGDealroom has a profile for this one. Try Dealroom →, formerly known as Sequoia Capital China, according to a person with direct knowledge of the matter. It brings total funding to $400 million across three rounds, first reported in mid-September 2026.
What's the endgame? Naive AI has since released its first large language model, Naive-N0.5-Flash, on September 27, 2026 — weeks after the funding was first reported. As planned, it opted for an open-weight release, similar to DeepSeek, MoonshotDealroom has a profile for this one. Try Dealroom →, and Alibaba, with the model's weights and inference code released under the MIT license for anyone to download and customise.
What's the approach? Rather than pretraining a model from scratch, Naive AI builds on an existing Chinese open-weight model, modifying its structure and refining it through reinforcement learning. The firm, which has fewer than 100 employees, is betting there is room for developers that excel at later-stage techniques like midtraining and posttraining. It is also researching recursive self-improvement, a focus shared by OpenAI, Anthropic, and GoogleDealroom has a profile for this one. Try Dealroom →.
Why now? Dai's rapid funding reflects a wider trend in China, where academic researchers have launched AI labs and raised money this year. Most of these "neolabs" focus on world models; Naive AI is one of few building LLMs.
Who's behind it? Dai is a prominent figure in Chinese AI academia, with a decade of work in deep learning, computer vision, foundation models, and agentic AI. He previously built InternVL, a widely adopted open-weight model, and worked on MiroThinker at startup MiroMindDealroom has a profile for this one. Try Dealroom → before leaving in January over disagreements with co-founders who wanted to relocate the team out of China. For now, Naive AI's website shows only a slogan — "100x Intelligence for the pioneers" — and no company details.
The signal: Building on existing open-weight models echoes a US pattern: Mira Murati's Thinking Machines Lab said its first model "largely follows" DeepSeek's V3, and Cursor used a model from China's Z.ai for its coding model. Even in a race dominated by tech giants and major labs, well-funded newcomers are betting that post-pretraining expertise can carve out a place.
Read more: The Information
Image credit: Generated with Gemini
