A Tsinghua Professor’s Stealth LLM Startup Hits $1.4 Billion Valuation

A secretive Chinese AI model startup founded in February by a Tsinghua University professor is now valued at more than $1.4 billion, after raising $400 million in three funding rounds from investors including Tencent, according to a person with direct knowledge of the matter.

Beijing-based Naive AI is preparing to release its first large language model as early as this month, joining the country’s fierce LLM competition as a newcomer, according to the person. The startup plans to release it as an open-weight model, similar to offerings from DeepSeek, Moonshot and Alibaba that people can download for free and customize to their liking.

Naive AI is led by Jifeng Dai, a renowned professor in China’s AI academia whose work for the past decade spans deep learning, computer vision, foundation models and agentic AI. For the past seven months, the startup has kept a very low profile, and its website only shows a slogan, “100x Intelligence for the pioneers,” and no information on the company. It just created an X account this month but hasn’t posted anything.

Naive AI’s investors include Tencent, IDG Capital, MPCi and HSG, formerly known as Sequoia Capital China, according to the person. The startup raised $100 million in its first funding round, $180 million in the second one and $120 million in the third one, which closed recently at a post-money valuation of $1.42 billion, the person said.

The rapid funding of Dai’s startup is part of a bigger trend in China, where many academic researchers and other entrepreneurs have launched their own AI labs and raised money this year. While most of those Chinese “neolabs” focus on world models that simulate the physical world, Naive AI is one of few that are developing LLMs.

Instead of creating its model from scratch in a costly process known as pretraining, Naive AI is building its LLM—which is also called Naive—based on an existing Chinese open-weight model, according to the person. (It couldn’t be learned which model Naive used.) Even though China’s LLM race is dominated by tech giants and major labs like DeepSeek, Moonshot and Z.ai, Dai is betting there is room for other developers that excel in techniques to enhance models’ capabilities in the later stages of development including midtraining and posttraining.

The startup, which has fewer than 100 employees, modifies the structure of a pretrained model and refines it through reinforcement learning and other processes to achieve better performance in various tasks, according to the person. The company is also researching recursive self-improvement, an AI model’s ability to automatically upgrade itself, which is a core focus of OpenAI, Anthropic, Google and many other developers.

Naive AI’s approach of utilizing existing open-weight models from other labs is similar to that of some U.S. neolabs. When Thinking Machines Lab, the startup led by former OpenAI executive Mira Murati, released its first LLM called Inkling in July, it said Inkling’s architecture design “largely follows” DeepSeek’s V3 open-weight model, which was released in late 2024. And Cursor, now owned by SpaceX, said it used a model from China’s Z.ai to develop its own coding model.

Prominent Chinese investors are backing Naive AI mostly because of their confidence in Dai’s expertise. A few years ago, he developed InternVL, an open-weight model that was widely adopted in the academic research community at the time. Last year, Dai joined a Chinese startup called MiroMind and worked on MiroThinker, an open-source AI agent for complex tasks such as multi-step web research and financial analysis. MiroThinker received positive feedback from developers and open-source communities on Reddit. But he left MiroMind in January after having disagreements with its co-founders who wanted to relocate the team out of China, according to Chinese media reports.

Among Chinese neolabs founded this year, Naive AI is one of the most well funded and highly valued ones. Another is Shanghai-based Pragmatik Labs, founded earlier this year by former Alibaba Group AI researcher Junyang Lin. Pragmatik, whose investors include Tencent, HSG and Gaorong Ventures, was valued at about $2 billion in its funding round a few months ago. Pragmatik is building digital AI agents for knowledge work as well as physical AI agents that enable robots to perform real-world tasks, according to its website.

Those new Chinese AI startups’ valuations are still tiny compared to U.S. neolabs founded in recent years by researchers who left OpenAI and other industry leaders. Thinking Machines Lab, for example, is in talks to raise between $5 billion and $6 billion at a pre-money valuation of at least $40 billion, The Information reported earlier this month. Chinese tech startups are typically valued far less than those in the U.S. due to more limited funding and exit options as well as the difficulty of monetization in the domestic market. Major Chinese labs such as Z.ai and Moonshot are generating revenue mainly from enterprise customers who pay for cloud-based access to their models through application programming interfaces

Juro Osawa is a reporter covering tech in Asia, from Alibaba and Tencent to startups. He previously worked for The Wall Street Journal. He is based in Hong Kong and can be found on Twitter at @JuroOsawa.