Harvey Introduces Harvey Tenet: A Kimi K3 Base Post-Trained with Fireworks for Long-Horizon Legal Agent Work
Harvey Tenet is a post-trained legal AI model built on the open-weight Kimi K3 base, fine-tuned via asynchronous reinforcement learning using Fireworks infrastructure Tenet achieves state-of-the-art on Harvey's LAB: Contracts benchmark and second place on the broader Legal Agent Benchmark (LAB), completing nearly twice as many held-out tasks as the base K3 model The model demonstrates strong zero-shot transfer to external benchmarks (Mercor's APEX Agents and Crosby's Redline Bench) without erodi
Analysis
TL;DR
- Harvey Tenet is a post-trained legal AI model built on the open-weight Kimi K3 base, fine-tuned via asynchronous reinforcement learning using Fireworks infrastructure
- Tenet achieves state-of-the-art on Harvey's LAB: Contracts benchmark and second place on the broader Legal Agent Benchmark (LAB), completing nearly twice as many held-out tasks as the base K3 model
- The model demonstrates strong zero-shot transfer to external benchmarks (Mercor's APEX Agents and Crosby's Redline Bench) without eroding foundational legal reasoning capabilities
- Training combined synthetic data, public legal corpora, and human expert data using GSPO optimization with rank-64 LoRA across ~1,750 sandboxed environments and over 10,000 rollouts per epoch on roughly 150 NVIDIA B300 GPUs
- Harvey also released three specialist post-trained models for M&A diligence, review tables, and firm knowledge management, each showing significant quality gains with reduced token consumption and cost
Why It Matters
Harvey Tenet represents a significant step toward making frontier-level legal AI accessible through open-weight models rather than proprietary black boxes, giving law firms a potential path to own and operate their own specialized models. The demonstration that agentic reinforcement learning on long-horizon legal tasks can transfer to unseen benchmarks while preserving foundational reasoning addresses a key concern in the industry about capability collapse during specialization. This approach could accelerate the adoption of autonomous legal AI agents in enterprise settings where data sovereignty and cost efficiency are critical procurement factors.
Technical Details
- Architecture and Base Model: Built on Moonshot AI's Kimi K3 open-weight model, with post-training applied via rank-64 LoRA adapters over the full K3 network using the GSPO (Generalized Soft Policy Optimization) algorithm
- Training Infrastructure: Co-built with Fireworks at the kernel level, featuring token-in-token-out processing and router replay to maintain numerical alignment of the large MoE (Mixture of Experts) architecture across training and inference
- Reinforcement Learning Setup: Asynchronous RL on sandboxed legal environments mimicking LAB tasks, with partner-style instructions (~50 words), client matters containing key and peripheral documents, and expert rubrics of roughly 50 atomic pass/fail criteria per task; single rollouts can exceed 1,000 turns
- Reward Function: Combines three components — fraction of rubric criteria satisfied, holistic count of legal issues solved, and an all-pass bonus; rollouts graded by LLM-as-a-judge (Kimi 2.6 selected through ablations)
- Specialist Sub-models: Three additional post-trained models routed by Tenet — an M&A diligence model using a Recursive Language Model harness with GLM-5.2 orchestrator (60.1% criteria pass on 80M-token tasks), a review table model with Applied Compute (10x cost reduction), and a firm knowledge model using Engram with Qwen3.8-27B (90% cost reduction per query)
Industry Insight
The emphasis on open-weight bases with proprietary post-training checkpoints reflects a growing industry tension between model accessibility and competitive differentiation; law firms should evaluate whether they can reproduce Harvey's recipe with their own data to build defensible specialized models rather than remaining dependent on vendor platforms. The demonstrated cost-quality tradeoff — achieving higher intelligence-per-token ratios through reward shaping that prefers shorter trajectories — suggests that agentic RL training can simultaneously improve capability and reduce operational costs, making enterprise deployment more economically viable at scale. The lack of public leaderboard presence for Tenet's benchmarks raises questions about independent verification that prospective buyers should scrutinize, particularly given Harvey's self-reported SOTA claims and the absence of third-party benchmarking on Vals or Artificial Analysis.
Disclaimer: The above content is generated by AI and is for reference only.