Harvey Introduces Harvey Tenet Legal Agent AI Model
Harvey launches Harvey Tenet, a post-trained legal agent model based on Kimi K3, built with Fireworks to excel on legal benchmarks.
Stock photo for illustration only, not from the actual event
- Harvey Tenet is post-trained on top of the open-weight Kimi K3 base model.
- Completes nearly twice as many tasks on Harvey's Legal Agent Benchmark (LAB).
- Co-developed kernel-level optimizations with Fireworks for training and inference.
- Underwent blind auditing by Mercor to verify real benchmark performance.
Harvey has officially unveiled Harvey Tenet, a specialized legal agent model built by post-training the Kimi K3 base model in collaboration with Fireworks. The new system is designed specifically to handle complex, long-horizon legal workflows that require multi-step reasoning and document analysis.
When evaluated against the base K3 model, Harvey Tenet successfully completes nearly twice as many held-out tasks on Harvey's Legal Agent Benchmark (LAB) and achieves a 20% increase in performance on LAB: Contracts, raising the overall all-pass rates by 9 and 2 percentage points respectively.
To address economic feasibility, Harvey co-optimized costs rather than trading them off against performance. By utilizing open-weight models, the team reduced the price per token, while reward shaping mechanisms encouraged shorter completion trajectories for equal quality, significantly lowering token consumption without sacrificing output standards.
"Against the base K3 model, Tenet completes almost twice as many held-out tasks on Harvey’s Legal Agent Benchmark (LAB) and 20% more on LAB: Contracts , lifting all-pass rate by 9 and 2 percentage points respectively."
Harvey
The training pipeline utilized asynchronous reinforcement learning inside sandboxed legal environments structured like LAB tasks. Grading was handled via LLM-as-a-judge, optimizing the policy with GSPO using a rank-64 LoRA over the entire K3 network across thousands of parallel environments and epochs.
Stock photo for illustration only, not from the actual event
The introduction of Harvey Tenet highlights a major trend in enterprise AI: adapting powerful general-purpose foundation models into highly regulated, domain-specific agents. Legal work requires strict adherence to factual rules, structured rubrics, and deep document comprehension, making specialized post-training architectures essential for real-world deployment.
Furthermore, Harvey demonstrated transparency by having Mercor run an independent blind test on APEX v1, alongside openly discussing divergences and benchmark nuances, setting a rigorous standard for legal AI evaluation.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment