NewsBite analysis · AI, Business, Data, Legal Industry Tech, Startups · New Product
MarkTechPost
A Kimi K3 Base Post-Trained with Fireworks for Long-Horizon Legal Agent Work
Harvey has unveiled Tenet, a post-trained model derived from the Kimi K3 base, aimed at long-horizon legal tasks. Utilizing synthetic, publicly available legal data alongside expert input, Tenet exhibits near double the task completion on Harvey’s benchmarks compared to its predecessor. The introduction of Tenet marks a notable step for law firms towards specialized legal intelligence models. The company aims to transition this research milestone to production while maintaining customer-centric approaches.
MonitorConfidence: HighMid term horizon
Published 23 Aug 2026, 18:40 UTC4 min read
Visit original article ↗Lead image — MarkTechPost
Summary
TL;DR
- Tenet achieves nearly double the completion rates on held-out legal tasks compared to the base K3 model.
- State-of-the-art results on LAB: Contracts and second place on LAB.
- Utilizes asynchronous reinforcement learning to enhance learning efficiency.
- Designed for various legal applications including M&A diligence and contract drafting.
- Harvey hopes to bridge research with production deliveries for law firms.
Interesting facts
- Tenet was trained using roughly 150 NVIDIA B300 GPUs over two months.
- The model's gains were observed in transfer learning contexts with external benchmarks.