Harvey launches tenet, a legal AI model built on kimi k3 with custom training

Harvey, an AI company backed by OpenAI, has unveiled Harvey Tenet, its first post-trained model, as a research preview. Tenet is based on the Kimi K3 model and was trained using asynchronous reinforcement learning through Fireworks, focusing on long-horizon legal tasks. The training data included synthetic data, public legal information, and input from human experts, without using customer data. Tenet outperformed the base K3 model on Harvey’s Legal Agent Benchmark (LAB), completing nearly twice as many tasks and improving the all-pass rate by 9 percentage points on LAB: Contracts and 2 percentage points on the general LAB benchmark. Harvey claims state-of-the-art performance on LAB: Contracts and second place on the overall LAB benchmark, using base-model scores from Vals. The model also showed strong transferability, improving performance on Mercor’s APE, and Crosby’s Redline Bench, which were not part of the training data. Tenet maintained performance on knowledge benchmarks such as LegalBench, CUAD, MAUD, and Scale’s PRBench, without compromising textbook legal reasoning. The training process involved sandboxed legal environments with partner-style instructions, client documents, and expert rubrics, with rollouts exceeding 1,000 turns per session. Evaluation was conducted by an LLM-as-a-judge, with rewards based on rubric criteria satisfaction, legal issue resolution, and an all-pass bonus. The model was optimized using GSPO with LoRA over the full K3 network, across ~1,750 environments and over 10,000 rollouts per epoch. Harvey also developed specialist models that Tenet can route to as tools or sub-agents. However, the release remains a research preview, with no model weights, model card, or API endpoint published. The company plans to move the work from research to production within its products over time. The release includes the training recipe, not the model itself. Harvey’s claims have been scrutinized by Reality Check, which found no contradictions in the published data, though some benchmarks remain unverified due to lack of public leaderboards. The score for Harvey Tenet was calculated as 78, based on self-reported metrics and no public leaderboard results.