Harvey Introduces Harvey Tenet: A Post-Trained Model for Long-Horizon Legal Work

On Harvey's Legal Agent Benchmark (LAB), Tenet completes almost twice as many held-out tasks as the base K3 model, and 20% more on LAB: Contracts, raising all-pass rates by 9 and 2 percentage points, respectively. Harvey reports state-of-the-art results on LAB: Contracts and second place on LAB, using base-model scores from Vals. The gains also transferred to Mercor's APEX Agents and Crosby's Redline Bench, tasks not seen during training, while performance on knowledge benchmarks such as LegalBench, CUAD, MAUD, and Scale's PRBench was maintained.

Training used asynchronous reinforcement learning in sandboxed legal environments modeled after LAB tasks. A single rollout can exceed 1,000 turns, with rollouts graded by LLM-as-a-judge. The reward combines rubric satisfaction, a count of legal issues solved, and an all-pass bonus. The policy was optimized with GSPO using a rank-64 LoRA over the full K3 network, running about 1,750 environments and more than 10,000 rollouts per epoch. Fireworks co-built trainer and rollout deployments at the kernel level with token-in-token-out and router replay to keep the large mixture-of-experts model aligned across training and inference. Training used roughly 150 NVIDIA B300 GPUs over two months, according to the report.

Harvey also post-trained three specialist models that Tenet can route to as tools or sub-agents. For M&A diligence, a post-trained GLM-5.2 orchestrator in a Recursive Language Model harness reached 60.1% criteria pass rate on LAB: Diligence, compared to 46.1% for the base orchestrator and 43.8% for the best baseline. A post-trained GLM-5.2 with Applied Compute improved answer quality by 3.6 points and citation quality by 12.1 points at about one-tenth the cost per cell. With Engram, a Qwen3.8-27B model studied 100 million tokens of client matters into 1 million tokens of structured knowledge, raising criteria pass rate by more than 15%, cutting tokens in completed trajectories by 58%, and reducing cost per query by roughly 90%.

Harvey Tenet is not yet deployable. It is a research preview, and Harvey has not published weights, a model card, or an API endpoint. The company said the work will move “from research to production” inside Harvey’s products over time. The stated goal is to build frontier legal intelligence on open-weight models and give law firms a path to own their own specialized models. Access runs through Harvey’s platform, which is sold to law firms, mid-sized firms, and in-house legal teams. Industries include legal services, corporate in-house legal, private equity, investment banking, and regulated sectors with high contract volume.