Harvey teamed with Fireworks AI to post‑train its Tenet model on a Kimi K3 base using asynchronous reinforcement learning. The updated model finishes almost twice as many Legal Agent Benchmark (LAB) tasks and raises the all‑pass success rate to 19.7%, compared with 10.8% for the original Kimi K3. Tenet also scores higher on related agentic benchmarks without losing legal knowledge and runs at roughly the same per‑task cost ($5.92 vs $5.62).
Why it matters
Law firms can get more complete legal drafts per dollar, reducing the time lawyers spend on repetitive document work.
