Harvey and Fireworks post-train Kimi K3 into Tenet for long legal tasks

Harvey introduced Tenet, its first model, post-trained with Fireworks from a Kimi K3 base using asynchronous reinforcement learning on the Fireworks Training API. The model targets long-horizon legal work, and the companies published benchmark results on August 26, 2026.

Benchmark gains and cost

On the Legal Agent Benchmark, Tenet reaches 19.7% all-pass, up from 10.8% for the base Kimi K3. On LAB Contracts, Tenet scores 11.3% all-pass versus 9.3% for the base model. Tenet completes almost twice as many held-out LAB tasks as the base Kimi K3, attains state-of-the-art performance on LAB Contracts, and ranks second overall on LAB.

On Mercor's Apex Agents (Corporate Law) benchmark, Tenet improves from 58.8% to 74.0% versus Kimi K3. On Crosby's Redline Bench, it improves from 49.3% to 55.5%. Harvey reports no significant regression on LegalBench, CUAD, MAUD, and Mercor's Apex-v1 benchmarks.

Cost per LAB task is effectively flat: $5.92 for Tenet versus $5.62 for base Kimi K3, despite higher throughput. Harvey attributes the flat cost to open-weight per-token pricing and the reward-shaping approach used in training.

Training and platform details

Training involved tens of thousands of long-horizon rollouts per checkpoint, each exceeding 50 turns and 100,000 generated tokens. Fireworks says its platform provides numeric alignment, sharing the same numerics between training and inference to avoid token clipping and reward collapse. The platform also ensures batch invariance: identical results regardless of other requests in the batch, with deterministic behavior at temperature zero.

A post-trained checkpoint can be promoted to production directly on the Fireworks platform without leaving the stack. Fireworks plans a forthcoming technical write-up detailing reward design, asynchronous RL setup, and lessons from large-scale long-horizon rollouts.

For inference engineers, the key takeaway is that long-horizon agentic tasks can be post-trained without blowing up per-task cost, provided the training and serving stack share numerics and maintain deterministic batching. The flat cost suggests that reward shaping and open-weight pricing can offset the added compute of extended rollouts, making such models viable for production legal workflows.