Most legal AI tools are still glorified search engines with a chat interface. Harvey is trying to build something different, and Harvey Tenet is the clearest sign yet of where that ambition is heading. Released on August 23, 2026 as a research preview, Tenet is Harvey’s first post-trained model. It takes Kimi K3 as a base and runs it through asynchronous reinforcement learning, trained specifically on long-horizon legal agent work. The result is a model that doesn’t just retrieve, it executes multi-step tasks across large document sets, the kind of work that previously required a junior associate and several hours.
The training corpus mixed synthetic data, publicly available legal data, and human expert annotation. Harvey says no customer data was used. The method matters here: Tenet was trained in sandboxed legal environments that mirror real tasks, structured as a partner-level instruction, a client matter with key and peripheral documents, and an expert rubric of roughly 50 atomic pass/fail criteria per task. A single rollout can exceed 1,000 turns. That’s not a question-answer format, that’s task completion under realistic conditions.
Against the base Kimi K3 model, Tenet completes nearly twice as many held-out tasks on Harvey’s Legal Agent Benchmark and 20% more on LAB: Contracts, raising all-pass rates by 9 and 2 percentage points respectively. Harvey reports state-of-the-art on LAB: Contracts and second place overall on LAB. But the more convincing result is what happened off-benchmark. Tenet improved substantially on Mercor’s APEX Agents and Crosby’s Redline Bench, neither of which appeared in training. That’s transfer, not overfitting, and it’s the kind of result that makes benchmark skeptics pay attention.
Equally important, agentic training didn’t erode baseline legal reasoning. Performance held on LegalBench, CUAD, MAUD, and Scale’s PRBench. That’s a real concern with RL fine-tuning on narrow tasks, and Harvey appears to have avoided the typical regression.
Cost was treated as a constraint, not an afterthought. Reward shaping penalized unnecessarily long trajectories, so the model learned to reach the right answer efficiently. Open weights from Kimi K3 lower the per-token cost floor. The combination means Harvey claims quality gains at stable cost, which is exactly the argument enterprise legal teams need to justify adoption internally.
Harvey also post-trained three specialist models that Tenet can call as sub-agents:
- M&A diligence: A Recursive Language Model harness where a root agent holds a full dataroom in a REPL and delegates to sub-agents. Post-training pushed criteria pass rate from 43.8% to 60.1% on tasks that can span up to 80 million tokens.
- Review Table: A post-trained GLM-5.2 that improved answer quality by 3.6 points and citation quality by 12.1 points at roughly one-tenth the cost per cell, with learned abstention when a question doesn’t apply.
- Firm knowledge: A Qwen3.8-27B model that compressed roughly 100 million tokens of client matters into structured memory. Criteria pass rate rose more than 15%, tokens per completed trajectory fell 58%, and cost per query dropped around 90%.
So what’s the catch? Tenet is not available yet. There are no public weights, no model card, no API. What Harvey has released is the method and the benchmark results. The model itself will move into Harvey’s products over time, accessible through their enterprise platform sold to law firms, in-house legal teams, and corporate legal departments in sectors like private equity, insurance, and financial services. A research lab with the right RL stack and around 150 NVIDIA B300 GPUs could replicate the approach, but few have that setup, and even fewer have Harvey’s proprietary legal training environments.
For context, Harvey is operating in a space where Thomson Reuters, Lexis, and a handful of well-funded startups like Ironclad and Spellbook are all pushing toward agentic legal workflows. What separates Tenet from most of those efforts is the open-weight foundation combined with domain-specific post-training, a path that gives law firms a credible route toward owning and customizing their own models. That’s a meaningful shift in the enterprise AI conversation, where data ownership and model transparency are increasingly non-negotiable for large legal buyers.
The research preview is a proof of concept that Harvey can train models, not just deploy them. Whether that translates into a production advantage depends on how fast they move from recipe to product.
