01.AI stops pre-training new large models from scratch

01.AI, the Chinese lab Kai-Fu Lee founded in 2023 to compete at the frontier of large language models, stopped pretraining new large foundation models from scratch and repositioned itself as a vendor of customised enterprise AI tools, Lee said in interviews through early 2025. Rather than continuing to build its own frontier models, the roughly 200-person company shifted to fine-tuning other labs’ open-weight releases — principally DeepSeek’s and Alibaba’s Qwen — and wrapping them in data-integration software for corporate clients in finance, gaming and legal sectors, a product Lee described as “the Palantir of China.”

Lee attributed the pivot to the economics DeepSeek’s January 2025 releases had exposed: he said DeepSeek’s training costs were a small fraction of what OpenAI had spent, and that “only a handful of companies with essentially bottomless balance sheets” could still justify building frontier models from scratch. 01.AI also entered a joint laboratory arrangement with its major investor Alibaba to continue large-model research jointly rather than alone.

The retreat made 01.AI one of the more explicit cases of a well-funded Chinese lab abandoning frontier pretraining in favour of the application layer once cheap, capable open-weight models from DeepSeek and Qwen removed the commercial case for training a comparable model independently — a dynamic that reshaped strategy across several second-tier Chinese AI companies during 2025.