From the source
Introducing Harvey Tenet, our first post-trained open-weight model, with promising initial results for legal AI performance and cost-efficiency.
Over the past six months, Harvey’s research agenda has focused on two goals: 1.
Building frontier legal intelligence using open-weight models; and 2.
Creating systems to allow law firms to build their own specialized models and own their intelligence.
Today, we’re sharing an update on that research effort, including initial results from our first post-trained model, which we’re calling Harvey Tenet.
Harvey Tenet is a Kimi K3 base that we post-trained together with Fireworks research for long-horizon legal work.
In addition, it incorporates harness improvements to make training and task execution more effective.
Our initial work shows promising results for both performance and cost-efficiency.
Quality Our goal in training was to improve the model's ability to perform long-horizon, agentic legal tasks.
Post-training on a combined corpus of synthetic data, publicly-available legal data, and human expert data substantially improves model performance on LAB hold-out tasks.
…
