U.S. AI startup shifts to chinese open-source model for legal AI development

A U.S. artificial intelligence startup that received investment from OpenAI has announced it is building its first internal model using the open-weight Kimi K3 from Chinese AI lab Moonshot AI. This move highlights a growing trend among Western tech companies to adopt Chinese open-source models as AI development costs rise. The legal tech company Harvey, based in San Francisco, revealed that its new model, Harvey Tenet, was trained on the Kimi K3. Harvey, which also has backing from Sequoia Capital and Andreessen Horowitz, previously focused on customizing closed-source models from U.S. AI giants like Anthropic, OpenAI, and Google. AI policy researcher Simon Hedlin noted that Harvey’s shift demonstrates the value of open-weight models, allowing developers to fine-tune them with industry-specific data to improve accuracy and reduce costs. Hedlin added that the U.S. is falling behind in the development of high-capability open-weight models. Harvey, founded in the summer of 2022, recently secured a valuation of $11 billion in a March funding round. In a blog post, Harvey stated that its research over the past six months focused on leveraging open-weight models to create 'frontier legal intelligence' for law firms. The company claims that Harvey Tenet outperformed both its underlying base model and U.S. advanced AI systems like Fable 5 and GPT-5.6 Sol in complex, long-term legal tasks. Harvey also highlighted cost efficiency, noting that open-weight models typically have lower per-token costs, and the company further reduced expenses by minimizing token usage during inference. The training of Harvey Tenet took two months and used approximately 150 NVIDIA B300 GPUs. Meanwhile, according to a report by The Information, AT&T’s vice president Mark Austin stated that open-source models, including NVIDIA’s Nemotron series, now handle about 40% of AI queries initiated by AT&T employees.