Legal artificial intelligence startup Harvey has unveiled its first in-house large language model for legal work, "Harvey Tenet". Harvey has grown a legal software business valued at $11 billion by relying on external models such as OpenAI and Anthropic. With the launch of its own model, it aims to secure both its cost structure and control of its business.

On Aug. 18, local time, Business Insider reported that Harvey set a goal of enabling lawyers to handle work that used to take hours to days at lower cost through Tenet. Harvey has so far provided its service by combining multiple external models. Its costs rose quickly as usage increased because it had to pay providers each time lawyers called those models. If it secures an in-house model with a certain level of performance, it can shift more work to an internal engine to reduce external model fees and open a path to improve profitability without raising customer charges.

The launch also coincides with a trend of large general-purpose AI companies moving directly into the legal market. Anthropic is targeting lawyers with plugins for document review and drafting, and OpenAI has moved to expand its legal business by recruiting Jason Boehmig (제이슨 뵈미그), the founder of Ironclad. Google and Meta are also being mentioned as possible followers. This trend raises uncomfortable questions for Harvey: what happens if the companies supplying its models also go after its customers, and how long it would take one of them to catch up. Having its own model could be a way for Harvey to control costs and its business direction.

Gabe Pereyra (게이브 페레이라), a Harvey co-founder, cited quality as well as cost as a reason for developing its own model. He co-founded Harvey with Winston Weinberg (윈스턴 와인버그), a former researcher at Google DeepMind. Pereyra said Harvey already allocates work across different models depending on the nature of tasks, and Tenet will be another option within that mix for customers. He added the model was refined around work that customers actually consider important.

Harvey also focused on creating legal reasoning data itself. It needed to create such data first in order to train the model in lawyers' ways of thinking. The company hired full-time and contract lawyers to design hypothetical disputes and case files and to assess how well the model reasons through them. It used firms such as Mercor and Snorkel to source contract workers.

Based on the data it secured, Harvey trained Tenet on a low-cost open-source model, "Kimi K3", from Chinese startup Moonshot. Kimi K3 has drawn industry-wide attention for its performance and price competitiveness since its launch in July. Harvey added its own legal data to tailor the model for legal work.

Tenet is part of a product revamp Harvey calls "Harvey 2". Anique Drumright (아니크 드럼라이트), Harvey's chief product officer, said it is also adding a new memory feature that lets an agent carry instructions across multiple tasks by saving how users work and what they prefer.

Harvey said it will soon release research comparing Tenet's performance on legal work with other models. But its own benchmark results need to be interpreted with some caution. Model developers use such tests to identify weaknesses and then retrain to boost scores, which is similar to helping write an answer key and then taking the same exam.

Tenet has not yet been applied to Harvey's service, and the company did not disclose a specific timing for adoption. Pereyra did not reveal the names of law firms running trial deployments of Tenet, but spoke candidly about its longer-term goal. He said he ultimately wants Tenet to become a "building block" for law firms to train their own models.

Lawyers learn something from every case, such as how to negotiate difficult clauses or how a buyer will respond to specific conditions. Much of that know-how is buried in individual lawyers' heads or in old documents. Pereyra believes models trained inside law firms can extract that know-how and turn decades of accumulated experience into blueprints for agents that perform specific tasks.

Harvey's idea is that law firms are unlikely to build such models from scratch. If Harvey provides Tenet as a starting point, each firm can train its own version to match how its lawyers work. That would move Harvey beyond being a simple software supplier toward a business model closer to Big Four accounting and consulting firms that do more than sell technology by assembling it to fit customers' business characteristics.

That could create an odd paradox for Harvey, which has been assessed as being no more than a "ChatGPT wrapper". If Harvey actually realizes this plan, the wrapper part that was seen as merely a shell around external models could instead be reassessed as the most valuable layer in its overall business structure.