Harvey Introduces Harvey Tenet: A Kimi K3 Base Post-Trained with Fireworks for Long-Horizon Legal Agent Work
MarkTechPost Asif Razzaq
Harvey just showed off Tenet, a legal AI built on Kimi K3 and tuned for long, messy law work. It’s promising, but what ships now is the recipe, not a model you can actually use.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Harvey has unveiled Tenet, its first post-trained model, and is calling it a research preview. The model starts from Kimi K3, then gets trained with Fireworks using asynchronous reinforcement learning on long-horizon legal tasks. Harvey says the training mix used synthetic data, public legal data, and human expert data — but not customer data.
The numbers are the kind that will make legal AI vendors sit up. On Harvey’s own Legal Agent Benchmark, Tenet finishes almost twice as many held-out tasks as the base K3 model. On LAB: Contracts, it does 20% better, lifting all-pass rates by 9 and 2 percentage points respectively. Harvey says that’s enough to claim state-of-the-art on LAB: Contracts and second place on LAB, using base-model scores from Vals.
More interesting is what happened outside the homework set. Tenet also improved on Mercor’s APEX Agents and Crosby’s Redline Bench, two benchmarks it wasn’t trained on, while keeping its performance on LegalBench, CUAD, MAUD and Scale’s PRBench. That matters because it suggests the model learned something broader than benchmark tricks. Agentic legal work can still smell a lot like legal work.
Harvey also says the setup is designed to keep cost under control rather than treating quality as the only trophy. The company argues that open weights help lower price per token, while reward shaping that prefers shorter trajectories cuts down tokens used. Training was not cheap in compute terms — roughly 150 NVIDIA B300 GPUs over two months — but Harvey says quality improved without a cost hit.
Don't expect to deploy Tenet tomorrow. Harvey hasn’t released weights, a model card, or an API endpoint, and the August 20, 2026 announcement is explicitly a preview. Access runs through Harvey’s enterprise platform for law firms, mid-sized firms, and in-house legal teams. So for now, the big release is the method. The artifact stays in the vault.
My take — AI-written commentary, not fact-checked reporting
This is the familiar legal-AI move: open-weight base model up front, proprietary checkpoint behind the curtain. Very efficient, very enterprise, and just vague enough to keep the lawyers calm while everyone else waits for the actual weights. The real story is that the industry still rewards benchmark theater when the thing that matters most is whether a model can survive a dataroom without making a fool of itself.
Read more about this at: MarkTechPost