Harvey Launches Tenet, Its Proprietary Legal LLM

Harvey presented Tenet on Tuesday, its first proprietary model designed for legal work. However, the LLM is not yet operational, and the company has not provided a timeline, while announcing upcoming comparative research. Trained on Kimi K3 and synthetic legal data created with lawyers, Tenet aims to ultimately reduce reliance on third-party models and serve as a foundation for internal models within law firms. The deployment is part of "Harvey II," which includes a Memory function to apply business preferences across tasks.

Tenet is not in service, and Harvey does not provide a date

The new Tenet model is not yet operational in Harvey's products. The company does not specify when it will be. Harvey also states that it will soon publish research comparing Tenet to other models on legal tasks. Gabe Pereyra did not wish to name the law firms likely to test the system in the meantime.

Harvey wants to make Tenet a foundation for in-house models

Gabe Pereyra hopes that Tenet will serve as a base for law firms to train their own models. Harvey plans to provide this foundation and allow each firm to adapt it to its methods through learning from its internal practices. The company bets that few firms will launch such systems from scratch. According to Pereyra, a model trained within a firm could help unlock expertise that has been buried in practices and documents. Presented as paving the way for these uses, this direction would shift Harvey towards an activity closer to professional services firms, where technology is configured according to each client's specifics.

The model relies on Kimi K3 and synthetic legal cases

To develop Tenet, Harvey first gathered datasets designed to replicate the logic used by lawyers. The company engaged legal professionals, both internally and with partners like Mercor and Snorkel, to create fictional litigations and case files, then evaluate the models on their ability to handle these situations. This data was used to train a variant of Kimi K3, an economical open-source model from the Chinese startup Moonshot, which has garnered much attention since its release in July due to its performance-to-cost ratio. Tenet is part of a broader project called Harvey II. According to product lead Anique Drumright, this deployment includes a Memory function that records work preferences and allows agents to apply these instructions across different tasks.

Stated goal: reduce costs and add a specialized option

Harvey currently relies on models provided by players like OpenAI and Anthropic. Each call to these models via the platform incurs a payment to the provider, a cost item that can grow rapidly with usage. A high-performing proprietary model could allow for more queries to be handled internally and reduce these fees, potentially improving margins without raising prices for clients. Gabe Pereyra indicates that this approach addresses both cost and quality considerations. He notes that the company is already directing tasks to the models whose skills are best suited and asserts that Tenet will add a targeted option for the legal work that matters most to users. Tenet is thus presented as designed to handle more legal tasks over durations ranging from several hours to several days, with a targeted cost lower than that of third-party models. This strategy extends the evolution of a company valued at $11 billion, initially built on models from other providers, and now seeking to prove its ability to create one of its own.

The market also attracts Anthropic, OpenAI, Google, and Meta

The presentation of Tenet on Tuesday comes as holders of major generalist models target the legal market. Anthropic has attempted to attract practitioners with review and drafting plugins, and OpenAI has recruited Jason Boehmig, founder of Ironclad, to lead its approach to the sector. Google and Meta are expected to follow suit. In this competitive context, Harvey has announced Tenet as its first proprietary AI model.