Harvey builds its first in-house model on Moonshot's Kimi K3

Harvey, the OpenAI-backed legal-AI startup, said it had post-trained its first in-house model, named Tenet, on top of Moonshot AI’s open-weight Kimi K3 rather than a closed model from Anthropic, OpenAI or Google — the three it had previously relied on for its customised legal-work products. The company said it built Tenet in collaboration with Fireworks AI, applying specialised legal training data and human expert input on top of the Kimi K3 base.

Harvey reported that Tenet completed almost twice as many held-out tasks as base Kimi K3 on its internal LAB benchmark, and about 20% more on a contracts-focused variant, LAB Contracts, with all-pass rates up nine and two percentage points respectively. The company said Tenet reached state-of-the-art performance on LAB Contracts among the models it tested and placed second on the broader LAB benchmark — figures that are Harvey’s own, run on its own evaluation set, rather than independently reproduced.

The move was reported as a notable shift for a Western application-layer company: Harvey had previously built its products by customising closed, proprietary models from US labs, and its own $11 billion March 2026 funding round drew on that positioning. The South China Morning Post and other outlets framed the pivot as evidence that Chinese open-weight releases such as Kimi K3 — launched by Moonshot in July — had become credible starting points for specialised commercial products even at firms with deep ties to US labs, driven by lower inference cost and the ability to fine-tune weights directly rather than working through a closed API.