Harvey unveils Tenet, a legal model built on Kimi K3 with asynchronous reinforcement learning

Harvey announced Harvey Tenet, its first model that underwent post-training, as a research preview. Tenet is built on the base model Kimi K3 and received additional training in partnership with Fireworks using asynchronous reinforcement learning on long-term legal tasks. The training corpus combined synthetic data, publicly available legal information and human expert data, and Harvey emphasized that client data were not used. The model itself is a private checkpoint of Harvey; the base weights of K3 are open-weight, but Tenet has not been released as an independent model.

Against the K3 baseline, Tenet almost doubles held-out tasks in Harvey’s legal agent benchmark (LAB) and adds 20 percentage points in LAB: Contracts, with improvements of 9 and 2 percentage points in all-pass rate respectively. Harvey reports a state-of-the-art result in LAB: Contracts and second place in the overall LAB, based on Vals baselines. More notable is zero-shot transfer: Tenet improves significantly on Mercor’s APEX Agents for corporate law and Crosby’s Redline Bench, both absent from training, while preserving performance on knowledge benchmarks such as LegalBench, CUAD, MAUD and Scale’s PRBench. Agentic training did not impair basic legal reasoning ability.

Training was conducted in sandboxed environments constructed as LAB tasks: senior-partner-style instruction of average 50 words, a client file with key documents and scope, and an expert rubric with roughly 50 atomic pass/fail criteria per task (hundreds at the extreme). A single rollout can span 1,000 turns. Rollout ratings were performed by an LLM judge; but-chain evaluations led to selecting Kimi 2.6 as the judge. The reward combines a satisfied-criteria rate, a holistic count of resolved legal issues and an all-pass bonus. Policy was tuned in GSPO with LoRA of rank 64 across the full K3 network, eight task groups of eight rollouts per optimization step, over about 1,750 environments and more than 10,000 rollouts for convergence. Fireworks co-built the trainer and kernel-level rollout dispatcher with token-in-token-out and router replay to preserve numerical alignment of the large MoE between training and inference.

As of today Tenet is not deployable. Harvey has not published the weights, a model card or an API endpoint. Access will be through Harvey’s platform sold to large, mid-size law firms and in-house legal teams, Enterprise-only. A lab with an RL stack can reproduce the method; training consumed roughly 150 NVIDIA B300 processors over two months. What is provided now is the recipe, not the finished model.

Weights are open-weight and co-optimised rather than benchmarked. Open-weight reduces token price; a reward design that favours shorter trajectories at equal quality reduces token consumption. Harvey reports significant quality gains at stable cost. Targeted applications include due-diligence memos for mergers and acquisitions, contract drafting, review and redlining, structured extraction from up to 10,000 documents, and precedent search over the firm’s accumulated knowledge.

Harvey’s post-training team developed three capabilities separately (text truncated). In the absence of released weights, a model card, an API and peer-reviewed benchmarks, Tenet remains a research preview, an interesting step toward specialized legal models built on open-weight foundations, but not yet a production-ready offering.