Harvey is moving deeper into the technology underlying its legal AI platform with the introduction of Harvey Tenet, its first open-weight model post-trained specifically for legal work.

The research preview marks a shift from applying general-purpose AI models to legal workflows toward developing intelligence trained around the structure of legal work itself. Harvey says the longer-term goal is to help law firms build specialized models shaped by—and potentially owned by—their organizations.

Tenet is based on Kimi K3 and was post-trained with Fireworks AI for long-horizon legal work. Rather than training a foundation model from scratch, Harvey refined an existing open-weight model using realistic legal assignments that require an AI agent to search matter files, analyze documents and produce substantive work product.

The training data combined synthetic material, publicly available legal data and input from human experts. Harvey said no customer data was used.

Training environments were designed to resemble assignments from a law firm partner. Each included a short instruction, a collection of relevant and peripheral client documents, and an expert rubric describing the facts, conclusions, citations and recommendations expected in the final deliverable.

Harvey trained Tenet across approximately 1,750 legal task environments using roughly 150 NVIDIA B300 GPUs over two months.

According to Harvey, Tenet completed almost twice as many held-out tasks on its Legal Agent Benchmark as the underlying Kimi K3 model and completed 20 percent more tasks on the benchmark’s contracts component. Its all-pass rate increased by nine percentage points on the broader benchmark and two percentage points on contract tasks.

The company said Tenet achieved state-of-the-art performance on LAB Contracts and placed second on the broader Legal Agent Benchmark, which includes more than 1,200 tasks across 24 practice areas.

Harvey also reported gains on third-party evaluations covering corporate law, contract negotiation and legal reasoning. Because Tenet had not encountered those benchmarks during training, the company believes the results suggest that its improvements can transfer beyond the specific environments used to train it.

Cost was another focus. Harvey used its training process to reward efficient reasoning and tool use, encouraging the model to complete tasks with fewer tokens when comparable results were possible. The company said this allowed it to improve performance without increasing overall inference costs.

Alongside Tenet, Harvey disclosed separate research into specialized capabilities for M&A diligence, high-volume document review and firm knowledge.

In one M&A diligence experiment conducted with Baseten, a post-trained model working through a specialized agent framework increased its rubric pass rate from 46.1 percent to 60.1 percent. The tasks required agents to navigate data rooms containing as many as 80 million tokens of material.

For Review Table, Harvey worked with Applied Compute to improve structured data extraction across large document collections. The resulting model improved answer quality by 3.6 points and citation quality by 12.1 points compared with the strongest baselines, while operating at approximately one-tenth the cost per cell.

Research conducted with Engram explored how a model could internalize a firm’s accumulated knowledge. Harvey reported that the approach increased the criteria pass rate by more than 15 percent, reduced the tokens used in completed tasks by 58 percent and lowered the cost per query by 90 percent.

Those specialized capabilities were trained separately and could eventually operate as tools or sub-agents routed by a broader Harvey model system.

Harvey is positioning Tenet as an early research milestone rather than a finished, broadly deployed product. The company plans to expand its Legal Agent Benchmark across more jurisdictions, practice areas and workflows while increasing the computing resources available for future models.

The research also points toward a more customized future for legal AI. Instead of every firm relying on the same general-purpose model, Harvey envisions organizations developing specialized intelligence around their own precedents, workflows and institutional knowledge.

For Harvey, Tenet is the first significant step toward making that model-specific legal intelligence part of its platform.