Harvey Partners with FireworksAI to Optimize Open Source Legal AI
Harvey announced a collaboration with FireworksAI to train an open-source legal-specific model.
The hybrid legal agent primarily uses GLM 5.1 and consults Opus 4.7, averaging only 0.83 calls to the cutting-edge model per task, achieving an 18% full pass rate across 100 tasks, outperforming Opus's standalone 14%, with costs of only $368 compared to $954.
Post-training (SFT) improved the Kimi 2.6 model's full pass rate on the Legal Agent Benchmark from 11% to 15%, surpassing Opus's 14%, with costs of only $84, about 1/11 of Opus's.
Source: Public Information
ABAB AI Insight
Harvey, as a leading legal tech company, previously launched AI-assisted contract review and due diligence tools, focusing on vertical domain fine-tuning in 2024-2025. This collaboration with FireworksAI continues its path of lowering deployment barriers for law firms through open-source and hybrid architecture.
In terms of capital strategy, Harvey is investing resources in post-training of open-source models and routing system construction, significantly reducing inference costs and shifting legal AI from expensive cutting-edge models to scalable deployment solutions, helping law firm clients lower technology expenditures while providing FireworksAI with proprietary datasets in the legal field.
Similar to the hybrid agent practices of Cursor and Aider in the coding domain, legal AI is undergoing a transformation from being dominated by closed-source cutting-edge models to expanding through open-source and hybrid routing. Harvey aims to establish dual advantages in cost and performance within the vertical industry.
Essentially, this represents a transfer of pricing power: the open-source and hybrid architecture shifts the pricing power of legal AI from a few cutting-edge model suppliers to the vertical application layer. The mechanism lies in targeted post-training and intelligent routing, significantly enhancing cost-effectiveness, allowing small and medium-sized law firms to afford high-performance agents, thus accelerating the widespread penetration of AI in the legal industry.
ABAB News · Cognitive Law
Hybrid surpasses singular, routing is smarter than brute force.
The more precise the post-training, the closer the cost advantage to weapon-grade.
When open-source defeats closed-source in vertical tracks, the transfer of pricing power is complete.