Harvey Tenet: Kimi K3 Post-Trained Model Aims to Advance Long-Horizon Legal AI Agents

Legal AI startup Harvey has released a research preview of Tenet, its first post-trained model designed for long-horizon legal agent work. Built on the open-weight Kimi K3 base and co-developed with Fireworks, Tenet uses asynchronous reinforcement learning to handle extended legal tasks. The company states that the model nearly doubles held-out task completion on its internal Legal Agent Benchmark (LAB) compared to the base K3 model.

Tenet is not yet available as a product. Harvey has not published model weights, a model card, or an API endpoint. The current release is a technical recipe for building the model, not the artifact itself. Harvey says the work will move “from research to production” inside its products over time.

Benchmark claims and transfer results

Harvey reports that Tenet completes almost twice as many held-out tasks on LAB and 20% more on LAB: Contracts, raising the all-pass rate by 9 and 2 percentage points respectively. The company claims state-of-the-art on LAB: Contracts and second place on LAB, using base-model scores from Vals.

The more notable result is transfer. Tenet improved substantially on Mercor’s APEX Agents (corporate law) and Crosby’s Redline Bench, neither seen during training. Gains also held on standard legal knowledge benchmarks including LegalBench, CUAD, MAUD, and Scale’s PRBench. Harvey reports that agentic training did not erode textbook legal reasoning.

Training method and cost co-optimisation

Training used asynchronous reinforcement learning in sandboxed legal environments built like LAB tasks. Each environment includes a partner-style instruction averaging about 50 words, a client matter with key and peripheral documents, and an expert rubric of atomic pass/fail criteria. A single rollout can exceed 1,000 turns.

Rollouts are graded by LLM-as-a-judge, using Kimi 2.6 after ablations. The reward combines the fraction of rubric criteria satisfied, a holistic count of legal issues solved, and an all-pass bonus. The policy is optimised with GSPO using a rank-64 LoRA over the full K3 network, across approximately 1,750 environments and more than 10,000 rollouts per epoch. Fireworks co-built trainer and rollout deployments at the kernel level.

Cost is co-optimised rather than traded away. Open weights lower price per token; reward shaping that prefers shorter trajectories at equal quality lowers tokens consumed. Harvey reports significant quality gains at stable cost.

Verification and source limitations

All benchmark claims are self-reported. Harvey’s LAB launched deliberately without a public leaderboard. The company disclosed that Tenet ran in the standard public harness plus a finish tool carried over from training, while rival scores came from Vals. Harvey did not publish LAB scores with and without the tool, so the value of the harness is unquantified.

On APEX Agents, Tenet ran in Harvey’s internal bash harness while rivals used Mercor’s published numbers. Harvey discloses that the harness lifts bare K3 from 58.8% to 67.5%, within 0.1 point of leader Fable 5 at 67.4%, before any training.

Harvey had Mercor run APEX v1 blind, without disclosing runs, tasks or task-level scores back to Harvey. This provides the strongest verification in the post, though it evidences knowledge retention rather than agentic skill.

Datos clave

Aspect	Detail
Base model	Kimi K3 (open-weight)
Training partner	Fireworks
Training method	Asynchronous reinforcement learning with GSPO
Availability	Research preview only; no weights, model card or API released

Why this matters for Indian legal tech

For Indian law firms, legal departments, and AI developers in the legal space, Harvey Tenet represents a trend toward specialised, cost-efficient legal AI agents on open-weight models. The stated goal of giving law firms a path to own their own specialised models could eventually affect Indian legal technology procurement. However, the model is not yet available for testing, all benchmark claims are self-reported, and there is no timeline for production release. Indian readers should treat the claims as a research preview with significant source limitations.

Source: MarkTechPost – Harvey Introduces Harvey Tenet: A Kimi K3 Base Post-Trained with Fireworks for Long-Horizon Legal Agent Work