Harvey dropped a research preview today that’s less a product launch and more a proof of concept—and honestly, that’s the most interesting thing about it. Tenet, built on the open-weight Kimi K3 base and post-trained with Fireworks using asynchronous reinforcement learning, completes nearly twice as many held-out tasks on Harvey’s Legal Agent Benchmark compared to the base model. On LAB: Contracts, it’s a 20% jump. Those gains also transferred to benchmarks Harvey never trained on: Mercor’s APEX Agents and Crosby’s Redline Bench both saw substantial improvement, while knowledge benchmarks like LegalBench and CUAD held steady. That’s the part worth paying attention to. Agentic training didn’t erode the model’s underlying legal reasoning—it layered execution capability on top.
The training setup is absurdly detailed. Rollouts ran through sandboxed legal environments with partner-style instructions averaging 50 words, expert rubrics of roughly 50 atomic pass/fail criteria per task, and single rollouts exceeding 1,000 turns. Harvey used GSPO with a rank-64 LoRA across the full K3 network, roughly 1,750 environments, and over 10,000 rollouts per epoch—all on about 150 NVIDIA B300 GPUs over two months. LLM-as-a-judge grading settled on Kimi 2.6 after ablations. No customer data touched the training corpus, which Harvey is quick to emphasize given how sensitive law firm data is.
Harvey also built three specialist models that Tenet can route to as sub-agents. The M&A diligence specialist, post-trained on GLM-5.2 in a Recursive Language Model harness with Baseten, hit 60.1% criteria pass rate on LAB: Diligence—where no baseline cracked 43.8% and single tasks can span 80 million tokens. A review table specialist with Applied Compute improved citation quality by 12.1 points at roughly one-tenth the cost per cell. The firm knowledge model, a Qwen3.8-27B trained with Engram, compressed 100 million tokens of client matters into 1 million tokens of structured knowledge, cutting cost per query by roughly 90%.
But here’s the catch: Tenet isn’t deployable. No weights, no model card, no API endpoint. Harvey says the work will move from research to production inside its products “over time,” but what shipped today is the recipe, not the artifact. For a company selling to enterprise law firms, that’s a strategic choice—Harvey wants firms to see the path to owning specialized models without handing them the checkpoint. The open-weight base makes reproduction feasible for labs with RL infrastructure, but most firms won’t build it themselves. They’ll wait for Harvey to productize it. The question is whether the transfer results hold at scale, or if benchmarks tuned to Harvey’s own evaluation stack flatter what real legal work demands.
