Harvey previews Tenet, its first own legal model, post-trained from Kimi K3

Harvey/Harvey agents

Harvey and Fireworks AI used reinforcement learning on long legal tasks to post-train Moonshot's open-weight Kimi K3 into Tenet. It completes almost twice as many held-out LAB tasks as the base model at about the same cost. Harvey had relied on OpenAI, Anthropic and Google models until now.

• LAB all-pass rate 19.7% vs 10.8% for base Kimi K3.
• Tens of thousands of practice runs per checkpoint, each over 50 turns and 100,000 generated tokens.
• Gains carried over to tests it was not trained on: Mercor Apex Agents (corporate law) 58.8% to 74.0%.
• Cost $5.92 per LAB task vs $5.62 for the base model.