What happened
US legal AI company Harvey released Tenet, its first post-trained open-weight model, built on Chinese company Moonshot AI's Kimi K3. Training involved about 1,750 specialized legal work agent environments over roughly two months, and early tests showed Tenet completed nearly twice as many comprehensive legal tasks as the original model, with about a 20% increase in contract-related tasks.
Other overseas firms are also leveraging Chinese open-weight models. The company behind Cursor built its Composer 2 and 2.5 models on Kimi K2.5, while US startup Cognition and UK startup Cosine used Kimi models to enhance software engineering capabilities. Thinking Machines Lab, led by OpenAI's former CTO, adopted DeepSeek V3's MoE architecture for its Inkling model and used Kimi K2.5 to generate synthetic data for initial training.
These cases show Chinese open-weight models are entering global AI research and product development through multiple channels: as base models for further training, as sources of synthetic data, and as references for model architecture and training methods.
Why it matters
The rise of capable open-weight models from China gives AI companies an alternative to closed-source models, particularly for vertical applications where task execution in real-world scenarios is crucial. Post-training in realistic environments helps models learn skills like information search, tool calling, and error correction, which general question-answering models often lack.
Open-weight models also offer practical advantages in cost and control. Companies can deploy them independently, continue training, and reduce reliance on external APIs, which may lower inference costs and increase flexibility. This suggests Chinese open-weight models are becoming a more integral part of the global AI supply chain, extending influence from product exports to technology output.
Key facts
Harvey's Tenet model is based on Moonshot AI's Kimi K3 and was trained with about 1,750 legal work agent environments over roughly two months.
Cursor's Composer 2 and 2.5, Cognition's software engineering models, and Cosine's products are all built on Kimi K2.5 or Kimi models.
Thinking Machines Lab's Inkling uses DeepSeek V3's MoE architecture and Kimi K2.5-generated synthetic data for initial fine-tuning.
In early tests, most general models completed legal tasks at a rate below 10% under strict criteria, while Tenet completed nearly twice as many comprehensive legal tasks as the original Kimi K3.
What to watch next
Whether more overseas AI firms will adopt Chinese open-weight models as a base for their own products, especially as the models continue to advance and close the gap with leading closed-source systems.
How the growing role of Chinese open-weight models in global AI development might influence the competitive landscape, including potential shifts in AI supply chains and international technology collaboration.