A stealth-mode AI model startup, founded in February of this year, has surpassed a US$1.4 billion valuation after completing three funding rounds totalling US$400 million, according to sources familiar with the matter. The investor roster includes Chinese tech giant Tencent.

The company, Naive AI, is expected to debut its first large language model as early as this month, positioning itself as a fresh contender in China's fiercely competitive AI model arena. The startup plans to release its product as an open-weight model, allowing users to download it for free and customise it as needed.

The venture is spearheaded by Professor Dai Jifeng, a prominent figure in the academic AI community. Over the past decade, his research has spanned deep learning, computer vision, foundation models, and agentic AI. For the past seven months, the startup has maintained an extremely low profile, with its website naive.ai displaying only the tagline "100× intelligence for the pioneers" and no other corporate information. The company only recently created an X (formerly Twitter) account, @NaiveAIlab, which has yet to post any content.

According to insiders, investors in Naive AI include Tencent, IDG Capital, MPCi, and Sequoia China. The funding rounds comprised US$100 million in the first, US$180 million in the second, and a recent third round of US$120 million, culminating in a post-money valuation of US$1.42 billion.

Rather than pursuing the costly path of training a model from scratch, Naive AI is building its namesake model on top of an existing open-weight model already available in China, sources revealed. It remains unclear which specific base model the company is using. The startup, which employs fewer than 100 people, plans to modify the pre-trained model's architecture and employ techniques like reinforcement learning to enhance performance across various tasks.

The company is also investigating recursive self-improvement—the ability of an AI model to autonomously iterate and upgrade itself—a core research focus for the likes of OpenAI, Anthropic, Google, and numerous research institutions. Naive AI's strategy of leveraging open-weight models from other labs mirrors the approach taken by some emerging AI labs in the United States, such as Thinking Machines Lab, founded by former OpenAI executive Mira Murati, whose first model Inkling was described as "broadly following" the architecture of DeepSeek's V3 open-weight model released in late 2024.