OpenAI-backed legal firm Harvey rebuilds its model on China's open-weight Kimi K3
Harvey previously tuned closed models from Anthropic, OpenAI and Google; it now post-trains Moonshot AI's open base, citing inference cost and the ability to train on legal data.
Original cover art, generated for this story. THE VISSION does not republish third-party press imagery.
- Harvey, the San Francisco legal-technology company backed by OpenAI, has built a model called Harvey Tenet on Moonshot AI's open-weight Kimi K3 base.
- The company previously customised proprietary closed models from Anthropic, OpenAI and Google for legal work.
- It cites lower inference costs and the ability to post-train on specialised legal data for higher accuracy as the reasons for the switch.
- Harvey claims state-of-the-art performance on complex legal work with the result; no independent benchmark figures were published.
Harvey, a San Francisco legal-technology firm counting OpenAI among its backers, has rebuilt its core model on Chinese open weights. Harvey Tenet is post-trained on Moonshot AI's Kimi K3 base, replacing an approach that customised proprietary closed models from Anthropic, OpenAI and Google. The company says the switch was driven by lower inference costs and by the ability to post-train on specialised legal data for higher accuracy — something a closed API model does not permit in the same way.
Harvey describes the result as achieving state-of-the-art performance on complex legal work. No independent benchmark figures accompanied that claim, and the South China Morning Post's report did not include cost or performance numbers, so the scale of the improvement is the company's own characterisation rather than a measured result.
Simon Hedlin, an AI policy researcher quoted in the report, called it "a great example of open-weight models" allowing developers to specialise systems by post-training on industry data. The wider significance is the direction of the substitution: an American company with an OpenAI investment on its cap table concluded that a Chinese open-weight base served its economics better than the closed American models it had been building on.
This is the open-weights argument arriving in a regulated, high-stakes vertical rather than a hobbyist one, and the deciding factors were cost and the right to train on your own data — neither of which a frontier lab can concede without changing its product. For anyone weighing a closed API against an open base, the relevant question this raises is not capability parity but whether the ability to post-train on proprietary data is worth more than the gap at the frontier. Harvey's answer, with legal liability attached, was yes.