Legal AI company Harvey has released Tenet, its first proprietary model, built by post-training Kimi K3, the open-weight model from Chinese startup Moonshot, with help from Fireworks AI. This replaces Harvey's prior reliance on renting models from OpenAI, Anthropic and Google, notably reducing dependence on OpenAI, which is itself a Harvey investor. The move is framed as both a cost play (turning variable inference costs into a fixed cost) and a quality play, since Tenet is shaped for legal-specific reasoning using training data manufactured by hired lawyers who built mock disputes and case files. The piece also flags regulatory and licensing wrinkles: under the EU AI Act, a post-train likely falls short of the compute threshold that would make Harvey the 'provider' of a modified model, leaving obligations upstream with Moonshot; and Kimi K3's license requires a separate commercial agreement above $20mn of annual revenue from model-as-a-service operators, a threshold Harvey's $350mn+ run rate far exceeds.
Questions this post answers
What base model did Harvey use to build its own legal AI model Tenet?
Harvey post-trained Tenet on Kimi K3, the open-weight model released in July by Chinese startup Moonshot, with training infrastructure work done alongside Fireworks AI. This replaces Harvey's prior approach of routing all customer legal work through third-party models from OpenAI, Anthropic and Google. Developers tracking which base models power vertical AI products can follow moves like this on daily.dev.
Does Kimi K3's license let companies build commercial products on top of it for free?
Not above a certain scale. Kimi K3 permits derivative models but requires a separate agreement with Moonshot for model-as-a-service operators generating more than $20 million in revenue over any 12-month period. Harvey, which runs at over $350 million annualised, exceeds that threshold significantly, making its licensing terms with Moonshot a notable open question. Anyone weighing open-weight license terms before shipping a commercial product can find this kind of detail on daily.dev.
Does a company that post-trains an open-weight model become responsible for its EU AI Act obligations?
Only if the modification is significant, which the European Commission's guidelines peg indicatively at around a third of the original training compute. A typical post-training run falls well below that threshold, meaning most compliance obligations remain with the original model provider rather than the company doing the fine-tuning, even for regulated uses like legal services. Teams navigating AI Act compliance for fine-tuned models can keep up with rulings and guidance via daily.dev.