Harvey Tenet: Post-Trained Kimi K3 Doubles Legal Agent Task Completion

Harvey has released Harvey Tenet as a research preview—its first publicly described post-trained model. Tenet takes Kimi K3 as a base and applies asynchronous reinforcement learning, executed in partnership with Fireworks at the kernel level, across sandboxed legal work environments drawn from Harvey's Legal Agent Benchmark (LAB) task structure. The training corpus combined synthetic data, publicly available legal data, and human expert data; Harvey states no customer data was used. The result is a domain-specialised checkpoint, not a general-purpose model: no weights, model card, or API endpoint have been published as of the August 20, 2026 preview date.

Against the base K3 model, Tenet completes almost twice as many held-out LAB tasks and 20% more on LAB: Contracts, raising all-pass rate by 9 and 2 percentage points respectively. Harvey reports state-of-the-art on LAB: Contracts and second place on LAB overall, with base-model scores sourced from Vals. The more important result is transfer: Tenet improved substantially on Mercor's APEX Agents (corporate law) and Crosby's Redline Bench—neither seen during training—while holding performance on LegalBench, CUAD, MAUD, and Scale's PRBench. Agentic training did not erode textbook legal reasoning, a meaningful distinction given the unresolved questions about how to verify gated capabilities across deployment contexts.

Training Architecture

Each environment mirrors a LAB task. A partner-style instruction averaging roughly 50 words opens the episode—no output specification. The agent receives a closed-universe matter file set mixing key and peripheral documents, with legal issues distributed across multiple files. A sandboxed workspace exposes tools for matter search, document reading, work product drafting, and disk writes. A single rollout can exceed 1,000 turns before the agent writes its final deliverables and ends the episode.

Scoring uses LLM-as-a-judge; ablations settled on Kimi 2.6. Each task carries an expert rubric of atomic pass/fail criteria—roughly 50 per task, with hundreds at the upper extreme. The reward function combines the fraction of rubric criteria satisfied, a holistic count of legal issues resolved, and an all-pass bonus. Policy optimisation uses GSPO with a rank-64 LoRA over the full K3 network, processing eight task groups of eight rollouts per optimiser step, across approximately 1,750 environments and more than 10,000 rollouts per epoch. A length term in the policy update penalises unnecessarily verbose deliverables, building cost co-optimisation into training rather than layering it on afterward.

Fireworks co-built trainer and rollout deployments at the kernel level, implementing token-in-token-out and router replay to keep the large MoE numerically consistent between training and inference. Rollouts and the trainer run asynchronously; after each optimiser step, new weights load into live deployments without stopping generation. Total compute: approximately 150 NVIDIA B300 GPUs over two months.

Benchmark Results

Three Specialist Sub-Models

Harvey also post-trained three capability-specific models designed to be routed to as tools or sub-agents:

M&A Diligence (with Baseten): LAB: Diligence tasks can traverse up to 80M tokens of document context. No baseline passed more than 43.8% of rubric criteria. Harvey adopted a Recursive Language Model harness in which a root agent holds the dataroom in a REPL and delegates to sub-agents. A GLM-5.2 orchestrator in that harness reached 46.1%; post-training it via self-distillation within the harness reached 60.1%. The training fix addressed the base model's pattern of under-delegating dataroom review.

Review Table (with Applied Compute): A post-trained GLM-5.2 improved answer quality by 3.6 points and citation quality by 12.1 points against the strongest baseline, at approximately one-tenth the cost per cell. Trained inside the production Review Table harness across up to 10,000 documents, the model learned to abstain when a question does not apply.

Firm Knowledge (with Engram): A Qwen3.8-27B model ingests approximately 100M tokens of client matter content, consolidates it into 1M tokens of structured knowledge, and internalises that knowledge parametrically through distillation and RL over self-generated data. Criteria pass rate rose more than 15%, task completion rose approximately 10%, tokens in completed trajectories fell 58%, and cost per query fell roughly 90%. Harvey reports an intelligence-per-token score of 190.8 on LAB: Firm Knowledge for the post-trained Qwen3.8-27B, against 129.3 for the best frontier configuration and 37.2 for equivalently sized models.

Availability

Tenet remains a research preview. Access runs through Harvey's enterprise platform, sold to law firms, mid-sized firms, and in-house legal teams, with stated industries extending to private equity, investment banking, insurance, financial services, healthcare, and energy. Harvey explicitly frames the open-weight base model as a path for law firms to eventually own their own specialised checkpoints.

The recipe Harvey is publishing—RL over expert-rubric environments, cost-aware reward shaping, asynchronous weight reloading into live inference—is reproducible by any organisation with an RL stack and the roughly 150 B300 GPU-months of compute. This pattern, where pipeline architecture rather than raw model scale drives meaningful capability gains, has become the dominant story of 2026's most credible AI deployments. Harvey's move to publish training specifics while withholding weights suggests the company expects the gap between method and artifact to close on its own timeline.