Harvey unveils first in-house legal AI model
▼ Summary
– Harvey launched Tenet, its first proprietary legal model, post-trained on Moonshot’s open-weight Kimi K3 with Fireworks AI, replacing reliance on models from OpenAI, Anthropic, and Google.
– OpenAI is a Harvey investor, so Tenet reduces dependence on a backer and rival, converting variable model-call costs into fixed ones.
– Tenet was trained on mock disputes and case files created by hired lawyers, with grading of model reasoning, to tailor it for legal work.
– European deployment raises AI Act questions, but a post-train likely falls below the threshold for significant modification, shifting obligations to upstream providers starting in Beijing.
– Kimi K3’s license requires a separate Moonshot agreement for model-as-a-service operators earning over $20mn annually, a threshold Harvey exceeds at $350mn+, complicating its use.
Harvey has stopped renting the very thing its business depends on. The legal software company has launched Tenet, its first proprietary model, after years of routing customer work through models from OpenAI, Anthropic and Google. The move marks a strategic shift in how the company handles its core technology.
The base it chose is the detail worth stopping on. Tenet is post-trained on Kimi K3, the open-weight model released in July by the Chinese startup Moonshot, in work Harvey says it did with Fireworks AI. This selection carries significant implications given the company’s existing relationships.
Consider who is on the cap table. OpenAI is an investor in Harvey alongside Sequoia and Andreessen Horowitz, so a company valued at $15.5bn has built its flagship on a Chinese base to reduce its dependence on its own backer. The commercial logic is unsubtle. Every model call Harvey makes on a customer’s behalf is an invoice from a rival, and owning the engine turns a variable cost into a fixed one.
Quality is the other argument. Cofounder Gabe Pereyra, formerly of Google DeepMind, says Harvey already routes tasks to whichever model suits them, and Tenet adds an option shaped around legal work specifically. The training material was manufactured. Harvey hired lawyers, on staff and through firms including Mercor and Snorkel, to invent mock disputes and case files and then grade how well models reasoned through them.
Europe’s answer to Harvey is Swedish. Legora is chasing a $10bn valuation and selling to many of the same firms, which makes the choice of base model a competitive question as much as a technical one. For a European buyer the AI Act is the next question. A company that modifies someone else’s general purpose model becomes the provider of the modified version only when the change is significant, with the Commission’s guidelines using roughly a third of the original training compute as the indicative marker.
A post-train almost certainly falls well below that. Which means most of the obligations stay upstream, and a European law firm deploying Tenet ends up relying on a documentation chain that begins in Beijing. The licence has a threshold of its own. Kimi K3 permits derivative models, but requires a separate agreement with Moonshot for model-as-a-service operators above $20mn of revenue in any 12 months, and Harvey runs at more than $350mn annualised, in a market already shaped by free Chinese models.
The open-weight debate has mostly been about capability and price. Harvey has quietly turned it into a question about who your counterparty is when the work is covered by privilege.
(Source: The Next Web)
