Harvey has introduced the Harvey Tenet legal model, its first proprietary, in-house large language model built specifically for legal work, as the company moves to reduce its dependence on third-party AI providers and sharpen its defences against growing competition from general-purpose AI companies.
The announcement marks a strategic inflection point for a company that built an $11 billion business on top of models from OpenAI and Anthropic. Every time a lawyer uses one of those third-party models through Harvey’s platform, Harvey pays the model provider for that call. As usage scales, those fees accumulate rapidly. A capable in-house model could allow Harvey to route more work through its own engine, reducing the cost of inference without necessarily passing any price increase on to customers.
Why Harvey Tenet legal model matters for the competitive landscape
The timing is not incidental. Anthropic has been pursuing law firms with tools for document review and drafting, while OpenAI hired Ironclad founder Jason Boehmig to lead its own push into legal. Google and Meta are described as potentially not far behind. The central risk for Harvey is plain: its most important suppliers are also its most formidable prospective competitors.
Building Tenet is one response to that pressure. Gabe Pereyra, the former Google DeepMind researcher who co-founded Harvey with Winston Weinberg, said the company already routes different tasks to different models depending on their strengths. His argument is that Tenet gives customers another option in that mix, one shaped specifically around the work lawyers actually care about, rather than a general-purpose model adapted after the fact.
Training on legal reasoning: how Harvey built the dataset
Creating a legal-specific model requires legal-specific training data, and Harvey had to construct much of that from scratch. The company hired attorneys, both on staff and on contract through companies including Mercor and Snorkel, to devise mock disputes and case files. Those lawyers then graded the models on how well they reasoned through the material. The result was a dataset designed to teach a model to think the way a practising lawyer does, not just to retrieve relevant text.
Harvey then used that material to train a version of Kimi K3, a low-cost, open-source model from the Chinese startup Moonshot. Kimi K3 was released in July and has attracted considerable attention for its combination of capability and price.
Tenet is part of a broader set of updates Harvey is calling Harvey 2. Anique Drumright, Harvey’s chief product officer, said the rollout also includes a new Memory feature that lets users save preferences about how they work, allowing Harvey’s agents to carry those instructions consistently across tasks. The combination of a proprietary model and persistent user context is designed to make Harvey’s platform behave more like a trained colleague than a generic tool.
Harvey says it will publish research showing how Tenet performs against other models on legal tasks. That benchmark data deserves careful reading. Model developers routinely use evaluation tests to identify weaknesses, then train specifically to address them. Over time, a model’s own benchmark becomes a less reliable guide to real-world performance: the developer has, in effect, seen the exam before sitting it.
Tenet is not yet live inside Harvey’s platform, and the company has not said when it will be. Pereyra declined to name any law firms currently testing it.
The longer-term ambition is more structurally interesting than the immediate product release. Pereyra has described wanting Tenet to serve as a foundation that law firms could use to train their own models, built on how their specific lawyers work. Legal knowledge, he argues, is largely locked inside individual practitioners’ heads or buried in old documents. A firm-specific model could convert decades of institutional practice into a reusable, automatable resource.
If that vision is realised, Harvey’s business starts to resemble something closer to a professional services firm than a software vendor: not simply selling a product, but helping clients configure bespoke systems around their own working methods. The irony is that a company long characterised as a wrapper around other companies’ models could end up owning the most differentiated layer in the legal AI stack.