As the capabilities of China‘s open-weight models, such as Kimi and DeepSeek, continue to improve, more and more overseas artificial intelligence (AI) companies have begun post-training on the basis of China‘s open-weight models and develop autonomous models and products accordingly. The forms in which China‘s open-weight models participate in overseas AI research and development are also increasingly diverse, ranging from directly receiving further training as base models to generating synthetic data, providing model architecture and training method references. China‘s models are accelerating their integration into the global AI industry chain.
Harvey, a U.S. legal AI company, recently released its first post-training open weight model, Tenet. The model was developed based on the latest open weight model, Kimi K3, from the dark side of the Chinese moon.
The technical details released by Harvey show that the Tenet model‘s training system is highly suitable for real-world legal work scenarios, relying on approximately 1,750 specialized legal work agents environments to complete iterative optimization, and the entire set of specialized training work takes about 2 months. Early tests released by the company in May showed that, considering the “all standards pass” criterion, less than 10% of most general models in the market complete legal practice tasks completely, and there is a clear lack of utility in specialized scenarios. After specialized training, Tenet completed nearly twice as many comprehensive legal practice tasks as the original Kimi K3 model. In specialized tasks such as contract drafting, review, and negotiation, the number of completed tasks also increased by about 20%, while essentially retaining the general legal knowledge capabilities of the original model.
This is not an isolated case of China opening its weighting models to enter the overseas AI product R&D chain. Composer 2, a self-developed programming model released by the U.S. AI programming platform “Cursor” in March of this year, reported that the model is based on Kimi K2.5, relying on continuous pre-training based on code data to solidify the professional base capabilities, then through large-scale reinforcement learning to complete task optimization, achieving advancement to specialized intelligent models; Composer 2.5, released in May, is still based on Kimi K2.5, further expanding the pre-training scale and the number of reinforcement learning tasks. Cognition, an AI startup in the United States, and Cosine, an AI startup in the United Kingdom, each built their core AI software engineering products based on the Kimi model, enhancing software engineering capabilities through further large-scale reinforcement learning.
The impact of China‘s open-weight model also extends to the training data and model architecture. When developing its first open-weight model, Inkling, the Thinking Machine Lab founded by former CTO of OpenAI, Mira Muratti, did not directly adopt the Chinese model as the base model, but its official technical description said that Inkling‘s Hybrid Expert (MoE) architecture mainly follows the technical path of DeepSeek-V3. In the post-training launch phase, the company also used open-weight models, including Kimi K2.5, to generate synthetic data for the first round of supervised fine-tuning.
In the past, AI enterprises tended to rely heavily on closed-source models when pursuing cutting-edge model capabilities. As Chinese open-weight model capabilities such as Kimi and DeepSeek continue to improve, some models have entered the cutting edge, providing these enterprises with technological options beyond closed-source models. Moreover, as artificial intelligence evolved from question-and-answer tools to agents, the criteria for measuring model capabilities also extended from knowledge answers to complete execution of complex tasks. And the operational processes, judgment criteria, and error correction mechanisms required to complete complex tasks often exist in real-world business scenarios, assessment systems, and workflows. In this context, post-training for vertical AI places more emphasis on task execution ability. AI companies such as Cursor, Harvey, and Cognition are conducting large-scale reinforcement learning in task environments that are close to real business, allowing models to learn from information search, tool invocation, context management, and error correction, among other things.
In addition to model capabilities, open-weighted models have cost and control advantages. They can be self-deployable, continuously optimized, and continuously trained by enterprises, with the advantage of reduced unit inference costs, increased deployment flexibility, and reduced dependency on external application programming interfaces (APIs) in large-scale call scenarios.
As technological achievements such as weights, data, and architecture become more deeply integrated into the global AI R&D and industrial system, the role of China‘s open weights model in the global AI supply chain continues to strengthen. Its international impact is expanding further from product shipments to technology exports, showing a trend to become an important technology source for overseas AI companies to develop new models and new products.