Harvey launches Tenet, its first in-house legal model
Harvey spent years selling Big Law a stack built on other labs' models. Today it says it has one of its own — trained on mock disputes, cheap enough to leave on, and meant to become the starting weight for every firm's private fork.
Harvey introduced Harvey Tenet, its first in-house, proprietary model for legal work, as the intelligence layer inside a broader product refresh it is calling Harvey II. Cofounder Gabe Pereyra, a former Google DeepMind researcher who started the company with Winston Weinberg, told Business Insider that Harvey hired attorneys — on staff and through contractors including Mercor and Snorkel — to invent mock disputes and case files, then grade models on how they reasoned through them. The company used that material to post-train a version of Moonshot's Kimi K3, the low-cost open-source model released in July. Harvey says Tenet is "frontier-level on prominent legal benchmarks" and performs on par with the strongest general models at an open-source cost, which it claims makes it practical to run agents continuously across every matter. Those scores have not been independently published. Harvey says research is coming. Tenet is not live in the product yet, and Pereyra would not name any firm testing it or give a ship date.
The business reason is as important as the benchmark claim. Harvey built an $11 billion legal-software business on a mix of OpenAI and Anthropic models, and it pays those labs every time a lawyer hits send. A capable in-house engine lets it route more work off that meter — better margins without asking customers to pay more — and gives clients a model shaped around the work they actually bill. That urgency is not abstract. Anthropic has been selling lawyers plugins for review and drafting, and OpenAI hired Ironclad founder Jason Boehmig to run its legal push. We covered the lab side of that encroachment when Anthropic names first Head of Claude for Legal. The wrapper that used to sit comfortably on top of the frontier stack now has to assume the stack wants the client.
Harvey II is the product story wrapped around the model. Agents now open inside a matter or project — documents, parties, staffing, permissions, and ethical walls already in place — instead of starting from a blank prompt. A Memory feature keeps how a lawyer structures a summary, what they want cited, and how they write, and it follows them across Harvey, Word, and Outlook. Harvey says users can see, edit, or turn that memory off, and that it is never used to train models. The longer bet is more interesting than the memory toggle: Pereyra wants Tenet to become the starting weight for firm-specific models, so two firms on Harvey end up with different models because their own work shaped them. That would turn Harvey from a ChatGPT wrapper into something closer to a professional-services shop that sells the base model and then configures it. We tracked the valuation race last week — Legora seeks $10B+ valuation, doubling in four months — and Tenet is Harvey's answer to the question that race keeps raising: what is a legal-AI company worth if the labs decide to sell the same thing?
What to watch: whether Tenet actually ships into customer workflows this year, and whether any named firm is willing to train a private fork on top of it.
If Harvey can train a legal model on mock case files, should firms trust it with the real ones? Tell us in the comments.
