01.AI steps back from giant-model pre-training

Background

01.AI is an AI company led by Kai-Fu Lee that built and offered its own large language models, including the Yi models. Lee later said the Yi-Lightning model received strong market feedback in 2024^[3]. Its revenue did not come from model sales alone: Lee said about 20–30% of the previous year's revenue came from paid overseas productivity tools such as PopAi^[3].

What happened

In early January 2025, reports said that 01.AI was selling its pre-training team to Alibaba Cloud^[1][2]. The company said the reports were false^[1][2].

Lee then gave his own account of the arrangement. Alibaba would take on the training of very large models through a joint lab with 01.AI, and some 01.AI team members had joined that lab^[3]. 01.AI's own teams, he said, would concentrate on smaller, cost-effective models^[3]. His reasoning was about cost: only technology giants can bear the expense of training super-large models^[3].

The difference between the two descriptions matters. Handing a team over to another company and working with a partner in a joint lab, where some staff take part in the partner's giant-model training, are not the same arrangement, and it was the first that 01.AI disputed. But both accounts point in the same direction: the company was stepping back from carrying out the largest and most expensive training runs itself.

Key details

Lee said 01.AI would prioritise "faster, smaller, and more cost-effective models"^[3]. Wikipedia records that by March 2025 the company had stopped pre-training large language models, and was focusing instead on tailored business solutions built on DeepSeek's models^[4].

The commercial side of the shift followed quickly. On 17 March 2025, 01.AI launched the Wanzhi enterprise one-stop large-model platform, which offers enterprises deployment of DeepSeek models and also supports Qwen and Yi models^[5]. A platform that serves several model families, rather than only the company's own, is a different business from selling access to a self-trained model: it places the company in the role of integrating and deploying models for business customers.

Why it matters

Pre-training a very large language model is a capital-intensive activity. It calls for large clusters of accelerators, long training runs and repeated experiments, and the money is spent long before any revenue appears. When the head of an AI company says that only giants can carry that cost, he is describing a division of labour in which a smaller firm moves up the stack toward applications, deployment and services.

For 01.AI that meant betting on smaller models and on enterprise work: the Wanzhi platform, DeepSeek deployment, and paid productivity tools overseas^[3][5]. It also meant that a company known for training its own models would, at least for a period, build on models trained by others^[4]. The episode is a compact example of how the frontier of model training and the business of putting models to work can pull apart: the former concentrates among firms with the largest computing budgets, while the latter opens to companies that can integrate, adapt and support models for paying customers.