- Published on
OpenAI-backed U.S. AI Firm Switches to Chinese Open-Source Model, Builds Legal AI on Kimi K3
- Authors
- Name
- DeepHorus
San Francisco-based legal tech company Harvey has announced the launch of Harvey Tenet, a legal AI model built upon the Chinese open-source model Kimi K3, marking a significant shift in AI model development by Western tech firms. Previously, the company relied on closed-source models from U.S. AI giants like OpenAI, but has now opted for a more cost-effective Chinese open-source solution. This transition occurs against the backdrop of rising AI model development costs and highlights the value of open-source models in specialized industry applications.
Harvey, which serves large law firms and enterprises, was founded in the summer of 2022 and backed by investors such as Sequoia Capital and Andreessen Horowitz. In a funding round in March, Harvey achieved a valuation of $11 billion. The company stated that Harvey Tenet, developed over six months of research, was officially released on August 22, aiming to provide a more efficient and professional AI solution for the legal industry.
According to Harvey's announcement, Harvey Tenet is based on the Kimi K3 open-weight model developed by Moonshot AI and underwent post-training. Post-training involves further optimizing a general-purpose foundation model with specialized datasets to enhance its ability to perform specific tasks. Harvey Tenet's training lasted two months, utilizing approximately 150 Nvidia B300 GPUs, and has shown excellent performance in complex legal agent tasks.
Harvey noted in its official blog that after comprehensive legal dataset training, Harvey Tenet outperforms its underlying foundation model as well as U.S. frontier AI systems including Fable 5 and GPT-5.6 Sol on a range of long-horizon, complex legal tasks. The company emphasized that the model not only achieves industry-leading performance but also significantly improves cost efficiency. Open-weight models inherently have lower per-token costs, and Harvey further reduces usage costs by minimizing token consumption during inference.
AI policy researcher Simon Hedlin commented on this shift, calling Harvey's choice an excellent example of the value of open-weight models. He pointed out that developers can leverage such models for post-training based on industry-specific or corporate data, thereby increasing accuracy and reducing costs. Hedlin also noted that the U.S. has fallen behind in the development of frontier open-weight models and is still in the early stages of exploring the potential of high-capability open-weight models.
Meanwhile, other industries are also adopting open-source AI models. According to a report by The Information, Mark Austin, Vice President at U.S. telecom giant AT&T, stated that open-source models, including Nvidia's Nemotron series, handle about 40% of AI queries from AT&T employees. This demonstrates the widespread acceptance of open-source models in practical business applications.
Harvey's move may have broader implications for the AI industry, prompting more companies to consider using Chinese open-source models to reduce development costs. As AI technology continues to evolve, competition between open-source and closed-source models may intensify, especially in specialized vertical fields. This trend could also boost collaboration and influence of Chinese AI labs on a global scale.
Looking ahead, Harvey plans to continue optimizing the Harvey Tenet model and applying it to more legal scenarios. The company hopes that through open-weight models, law firms will be able to build and deploy their own specialized AI models, driving the intelligent transformation of the legal industry. This development may signal a greater focus on cost-effectiveness and industry customization in AI model development.
