Harvey Tenet: Legal AI Model Research Preview

Harvey Tenet: Legal AI Model Research Preview

Harvey Tenet is not a product you can buy or test today. It is a research preview, a carefully documented recipe for building a specialized legal AI that Harvey says will eventually power its platform. The company released the technical details on August 20, 2026, and the most striking part is not the model itself, but how it was trained and what it suggests about the future of legal tech.

Tenet starts with a Kimi K3 base model, an open-weight foundation. Harvey then post-trained it using asynchronous reinforcement learning. That's a method where the model learns by doing long, complex tasks and getting feedback, with the training data mixing synthetic examples, publicly available legal documents, and human expert input. No customer data was used. The company said so. But this is a deliberate choice, likely made to address privacy concerns that are acute in legal work, since clients in that field can't risk their confidential materials slipping into a model's training set, and they don't want to wonder where their secrets might end up.

The results are important. Against the base K3 model, Harvey Tenet completes almost twice as many held-out tasks on Harvey's own Legal Agent Benchmark (LAB). On the narrower LAB: Contracts test, it improves by 20%. The all-pass rate, meaning tasks fully completed without errors, rose by 9 percentage points on LAB and 2 points on Contracts. Harvey reports top performance on LAB: Contracts and finishes second overall on LAB, using base-model scores from Vals as a reference point.

Transfer Learning That Actually Works

The more interesting result emerges when Tenet confronts tasks it never saw during training. It improved substantially on Mercor's APEX Agents benchmark for corporate law and Crosby's Redline Bench. Those are external tests, not Harvey's own. So the model clearly learned general legal reasoning skills, not just how to game a specific benchmark, which is a far more meaningful signal for real-world deployment. It didn't sacrifice core knowledge either. Performance on LegalBench, CUAD, MAUD, and Scale's PRBench stayed stable. That's the kicker.

Here's the key point. It's a trap many specialized models fall into,they overfit to their training data, looking spectacular on internal tests but collapsing when they hit the messy, unpredictable real world. Tenet avoided that. The gains transferred, untrained, to benchmarks the model had never encountered. That's evidence of actual capability, not benchmark fitting. And that kind of transfer doesn't happen by accident.

Agentic training did not erode textbook legal reasoning. The model got better at doing legal work without forgetting what it already knew.

Cost Was Co-Optimized, Not Traded Away

Here's where the approach diverges from typical AI development. Most efforts zero in on quality alone, but Harvey Tenet's training used reward shaping that favors shorter paths when quality is equal, a subtle yet notable twist. In plain terms, the model is rewarded for being concise when accuracy is maintained. It's a cost killer. This lowers the number of tokens consumed per task, which directly cuts cost, and open weights already lower the price per token on top of that. Harvey said these choices together bring quality gains at a stable cost. So they didn't have to choose between better and cheaper. They got both.

The training setup was substantial, but not absurd. Roughly 150 NVIDIA B300 GPUs ran for two months. That's it. That's a large compute budget, yet it's within reach for a well-funded lab, since the company notes that a lab with a reinforcement learning stack could reproduce the method without needing any proprietary infrastructure. So the recipe is public, even if the final model isn't.

How the Training Actually Worked

The training environment mirrored a legal assignment. A partner-style instruction averaging about 50 words, paired with a client matter holding both key and peripheral documents, plus an expert rubric built from atomic pass/fail criteria. Those rubrics were dense. Roughly 50 criteria per task, with hundreds at the extreme, so a single rollout,one complete attempt at a task,could exceed 1,000 turns. That's a punishing chain. But sustaining that long sequence of actions is exactly what the model must do, and it's a far cry from a simple Q&A.

Rollouts were graded by an LLM-as-a-judge, with ablations settling on Kimi 2.6 as the grader. The reward function combined the fraction of rubric criteria met, a full count of legal issues solved, and an all-pass bonus. The policy was optimized with GSPO using a rank-64 LoRA over the full K3 network. Training ran eight task groups of eight rollouts per optimizer step, across roughly 1,750 environments and over 10,000 rollouts per epoch.

Fireworks is a partner in this effort. They co-built the trainer and rollout deployments at the kernel level, using token-in-token-out and router replay to keep a large mixture-of-experts model numerically aligned across training and inference. That's the kind of engineering detail that separates a working system from a paper concept. Harvey Tenet is a checkpoint, not a release. No weights published. No model card. No API endpoint. The base model is open-weight, but Tenet itself is Harvey's own property, and it's a distinction that can't be brushed aside. So don't mistake the architecture for the product.

The Specialist Models Show the Real Potential

Harvey didn't stop at the main model. Beyond it, he post-trained three specialist capabilities that Tenet can route to as tools or sub-agents, and those are where the largest performance deltas landed, marking the clearest gains across the entire system. But it's those three that really matter.

M&A Diligence: Handling a Torrent of Documents

On LAB: Diligence, a single task can traverse up to 80 million tokens. That is an enormous amount of text, far beyond what most models can process in one pass. No baseline passed more than 43.8% of the criteria. With Baseten, Harvey moved to a Recursive Language Model setup. A root agent holds the dataroom in a REPL, a read-eval-print loop, and delegates to sub-agents. A GLM-5.2 orchestrator alone reached 46.1%. Post-training in that setup via self-distillation jumped to 60.1%. That is a massive improvement on an already difficult task.

Review Table: Better Answers, Lower Cost

With Applied Compute, a post-trained GLM-5.2 improved answer quality by 3.6 points and citation quality by 12.1 points. This was achieved at roughly one-tenth the cost per cell. The model also learned to abstain when a question does not apply, a subtle but important behavior. Knowing when not to answer is a form of intelligence.

Firm Knowledge: Compressing a Firm's Memory

With Engram, a Qwen3.8-27B model studies roughly 100 million tokens of client matters. It compresses this into 1 million tokens of structured knowledge plus parametric memory. The criteria pass rate rose more than 15%. Tokens in completed trajectories fell 58%. Cost per query dropped roughly 90%. The intelligence-per-token metric hit 190.8 versus 129.3 for the best frontier configuration. This is the clearest example of the cost-quality co-optimization working in practice.

The specialist stack isn't just about cost savings. It's about making legal work tractable at scale. M&A diligence over datarooms, contract drafting, redlining, structured extraction across up to 10,000 documents, and precedent search over a firm's accumulated knowledge, these are the tasks that consume thousands of lawyer hours. So Harvey Tenet's research preview suggests a path to automating parts of them. That's the real shift.

Who Is This For?

The stated goal is twofold. First, build frontier legal intelligence on open-weight models. Second, give law firms a path to own their own specialized models. This is a direct appeal to firms that are wary of sending client data to a third-party API. The enterprise tier is the only option. Access runs through Harvey's platform, sold to law firms, mid-sized firms, and in-house legal teams. Industries include legal services, corporate in-house legal, private equity and investment banking for M&A diligence, plus regulated sectors where contract volume drives cost. Insurance, financial services, healthcare, and energy are all named.

Harvey Tenet is not deployable today. The company says the work will move "from research to production" inside its products over time. What ships today is the recipe, not the artifact. For a lab with the right stack, that recipe is a starting point. For everyone else, it is a signal of where legal AI is heading. The model is not the product. The method is.