Harvey's AI Handles M&A Diligence Across 10K Docs

Harvey announced a research preview on August 20, 2026, called Harvey Tenet, its first post-trained model built for legal work that spans thousands of documents at once, the kind of contract review or M&A diligence job that used to eat a junior associate's week. The starting point is Moonshot AI's Kimi K3, an open-weight base model. Harvey and Fireworks ran it through asynchronous reinforcement learning on long-horizon legal tasks, training on a mix of synthetic data, public legal filings, and human expert examples. No customer data went into the mix, according to Harvey.

The training run used roughly 150 Nvidia B300 GPUs over two months, and the payoff shows up on Harvey's own Legal Agent Benchmark, where Tenet clears far more held-out tasks than the untouched K3 base. Those gains reportedly carried over to outside benchmarks the model never trained on, including Mercor's APEX Agents and Crosby's Redline Bench. Nothing here ships to customers yet.

There's no public model card, no weights, no API. What Harvey has put out is the method, aimed at law firms that want their own specialized models rather than a rented API call.