Harvey introduced Harvey Tenet — the company’s first post-trained checkpoint for long-horizon legal work. It builds on Kimi K3, post-trained with Fireworks using asynchronous reinforcement learning. The corpus blended synthetic data, publicly available legal data, and human expert data; no customer data was used. Versus base K3, Tenet completes almost twice as many held-out tasks on LAB and 20% more on LAB: Contracts. Add transfer gains on external benchmarks — and you see the aim: frontier legal intelligence on open-weight models and a path for firms to own specialized models.
What is Harvey Tenet, and why does it matter?
Harvey Tenet is a post-trained Kimi K3 checkpoint focused on long legal processes. It was trained with asynchronous RL in partnership with Fireworks, targeting long agentic interactions. The goal is twofold: build frontier legal intelligence on open-weight foundations, and let law firms own their specialized models.
The training corpus mixed three sources: synthetic examples, publicly available legal data, and human expert contributions. The company explicitly states that no customer data was used. That boundary matters for trust and compliance.
Tenet targets “long horizons” — scenarios with hundreds or thousands of steps. Such patterns are common in due diligence, contract analysis, and precedent search. The strategy emphasizes guided trajectories, clear rubrics, and aligned LLM judges.
The result is substantial gains over base K3 on Harvey’s internal benchmarks. On LAB, Tenet completes nearly twice as many held-out tasks. On LAB: Contracts, it completes 20% more tasks, lifting all-pass rate by 9 and 2 points respectively.
Beyond that, improvements transfer to external suites without specific fine-tuning. The model advances on APEX Agents and Redline Bench while holding steady on knowledge tests. That signals learned behavior rather than metric fitting.
Is Tenet production-ready today?
Not yet. Harvey Tenet is a research preview announced on August 20, 2026. Weights, a model card, and an API endpoint are not available. K3 is open-weight, but Tenet itself is Harvey’s private checkpoint. The company says the work will move “from research to production” inside its products over time.
In effect, what shipped is a recipe, not an artifact. The team shares the approach: how to build agentic environments, shape rewards, and keep a large MoE numerically aligned between training and inference. This allows labs to attempt reproducing the method.
“What ships today is the recipe, not the artifact.”
The company tier is Enterprise. Access runs through Harvey’s platform, sold to law firms, mid-sized firms, and in-house legal teams. A lab with an RL stack could reproduce the method; training used roughly 150 NVIDIA B300 GPUs over two months.
Where is it relevant? Legal services, corporate in-house legal, private equity and investment banking (M&A diligence). Also regulated sectors with heavy contract volume: insurance, financial services, healthcare, energy. Sample applications include M&A diligence memos over datarooms, contract drafting, review and redlining, structured extraction across up to 10,000 documents, and precedent search over a firm’s knowledge.
What do the numbers show, and how does cost behave?
Against base K3, Tenet completes almost twice as many held-out tasks on the Legal Agent Benchmark. On LAB: Contracts, it completes 20% more tasks, 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, using base-model scores from Vals.
The more interesting result is transfer. Tenet improves substantially on Mercor’s APEX Agents (corporate law) and Crosby’s Redline Bench. Neither was seen during training. Meanwhile, the model holds performance on LegalBench, CUAD, MAUD, and Scale’s PRBench. Agentic training did not erode textbook legal reasoning.
Cost and quality are co-optimized rather than traded off. Open weights reduce price per token. Reward shaping that prefers shorter trajectories at equal quality lowers tokens consumed.
Harvey reports significant quality gains at stable cost. That matters for firms with massive contract sets and constrained budgets. Fewer tokens per episode mean smaller inference bills.
Is there a risk of “overfitting” to benchmarks? Results on APEX and Redline argue otherwise. Behavioral strategies transfer, while knowledge metrics do not slide. That is a strong sign of real agentic learning.
How was Tenet trained and kept aligned?
Training used asynchronous RL in sandboxed environments modeled on LAB tasks. Each episode includes a partner-style instruction of about 50 words, a client matter of key and peripheral documents, and an expert rubric of atomic pass/fail criteria. Typical tasks have around 50 criteria, and the extremes reach into the hundreds.
A single rollout can exceed 1,000 turns. Such long trajectories stress memory and policy stability. Evaluation is automated: an LLM-as-a-judge scores by the rubric, and ablations settled on Kimi 2.6 as the judge.
Reward combines three signals: the fraction of rubric criteria satisfied, a holistic count of legal issues solved, and an all-pass bonus. The policy is optimized with GSPO using a rank-64 LoRA over the full K3 network.
Each optimizer step processes eight task groups with eight rollouts each. Training spans roughly 1,750 environments and more than 10,000 rollouts per epoch. That breadth provides statistics across a wide variety of cases.
Fireworks co-built the trainer and rollout deployments at the kernel level. Token-in-token-out and router replay keep a large MoE numerically aligned between training and inference. Alignment reduces behavioral drift across phases.
Which specialist capabilities were trained separately?
The Harvey team also post-trained specialists that Tenet can route to as tools or sub-agents. The aim is to extend coverage on hard sub-tasks without overloading the base checkpoint.
M&A diligence. On LAB: Diligence, a single task can traverse up to 80M tokens. No baseline passed more than 43.8% of criteria. With Baseten, Harvey moved to a Recursive Language Model harness: a root agent holds the dataroom in a REPL and delegates to sub-agents. A GLM-5.2 orchestrator alone reached 46.1%; post-training it in that harness via self-distillation lifted results to 60.1%.
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 at roughly one-tenth the cost per cell. The model learned to abstain when a question does not apply.
Firm knowledge. With Engram, a Qwen3.8-27B model studies about 100M tokens of client matters into 1M tokens of structured knowledge plus parametric memory. Criteria pass rate rose by more than 15%, tokens in completed trajectories fell 58%, and cost per query dropped roughly 90% — 190.8 intelligence-per-token versus 129.3 for the best frontier configuration.
In aggregate, the largest deltas landed in the specialist stack: RLM diligence, Review Table, and firm memory. Tenet can call these tools as sub-agents, capturing gains where processes hurt the most.
Does that make the base checkpoint less important? No. It carries general agentic behavior and legal reasoning, while specialists cover narrow long tails. Together they form a flexible yet controlled working stack.
Based on original report.
