Product2 publishersIndependently confirmed2 min readPublished Updated
Harvey builds its flagship legal model on Chinese open weights, not its investor's API
Tenet is post-trained on Moonshot's Kimi K3, converting per-call payments to OpenAI, Anthropic and Google into a fixed training bill. The counterparty risk moved rather than disappeared.
The Product Desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Harvey has launched Tenet, its first proprietary legal model, post-trained on Kimi K3, the open-weight model from Chinese startup Moonshot.
- Tenet is built to displace OpenAI's models inside Harvey's product, and OpenAI is one of Harvey's investors, alongside Sequoia and Andreessen Horowitz.
- The training material was manufactured: lawyers on staff and via Mercor and Snorkel wrote mock disputes and case files, then graded model reasoning on them.
- Harvey says the post-training was done with Fireworks AI.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost Harvey stops paying per call to companies whose products compete with its own and starts paying up front for training and legal data work it controls.
- constraint Independence from OpenAI's meter runs through Moonshot's licence: Harvey's revenue sits more than 17 times above the threshold at which that agreement has to be negotiated.
- exposure European law firms handling privileged matters take on a supplier whose base-model documentation duties sit with a lab in Beijing rather than with Harvey.
- precedent Once a post-trained open base is deemed good enough for legal work, rivals such as Legora will be asked whose weights they are running as a sales question, not a technical one.
The base weights are the commodity in this arrangement. Anyone can download Kimi K3, released in July by Moonshot [2]. What cannot be downloaded is the grading: Harvey paid lawyers, some on staff and some through Mercor and Snorkel, to invent mock disputes and case files and then score how models reasoned through them [5]. That corpus exists because someone bought it, and it moves to whatever base looks best next year.
Cofounder Gabe Pereyra's description of the product, routing each task to whichever model suits it, is the doctrine rather than a hedge [9]. A system that already treats OpenAI, Anthropic and Google as interchangeable inputs [6] has defined its suppliers as swappable, and adding an owned option to that router is a procurement decision more than a research one.
The arithmetic makes the pressure legible. Harvey is valued at $15.5bn [8] on more than $350mn annualised [13], which puts the price at under roughly 44 times revenue [15]. A multiple in that range is a bet on margin expansion, and margin is hard to expand when the largest variable input is metered by firms selling into the same buying committee [4].
What the move does not remove is a counterparty. Kimi K3's licence permits derivative models but sends model-as-a-service operators above $20mn of revenue in any 12-month period to negotiate a separate agreement with Moonshot [12]. Whether a legal research product counts as model-as-a-service is the load-bearing definition here, and the source material does not settle it. If it does count, the negotiation is not a marginal one.
The regulatory result is stranger than the commercial one. Because a post-train sits well below the roughly one-third-of-original-training-compute marker in the Commission's guidelines, Harvey does not become the provider of the modified model, and most obligations stay upstream [11][17]. So Harvey gets the cost profile of ownership without the paperwork of authorship, and a European firm deploying Tenet inherits a documentation chain that begins in Beijing [17]. That is a manageable answer in a procurement questionnaire and a less comfortable one when the underlying work is privileged.
Legora, chasing a $10bn valuation and selling to many of the same firms [10], now has to answer the same question from the other side. In a market where Chinese weights are already free [18], the cheapest engine and the most awkward provenance are currently the same object.
What to watch
- Whether Harvey and Moonshot disclose a licence agreement, and whether Harvey's product is treated as model-as-a-service under it.
- Whether Harvey publishes evaluation results for Tenet against the OpenAI, Anthropic and Google models its router still calls.
- Whether Legora names its base model or commits publicly to frontier APIs as a differentiator with European buyers.