Harvey Previews Tenet: A Post-Trained Legal Agent Model Based on Kimi K3

Harvey has released Harvey Tenet, its first post-trained model, as a research preview. [1] The The model is based on Kimi K3 and was post-trained with Fireworks using asynchronous reinforcement learning on long-horizon legal work. [1] The training process, which required approximately 150 NVIDIA B300 GPUs over two months, utilized a mix of synthetic, public, and human expert data while excluding customer data. [1] The stated objectives for this release are to construct frontier legal intelligence using open-weight models and to provide law firms a pathway to own their specialized models. [1]

When compared to the base K3 model, Harvey reports that Tenet completes nearly twice as many held-out tasks on the Legal Agent Benchmark (LAB) and achieves a 20 percent increase on LAB: Contracts. [1] The model secured state-of-the-art results on LAB: Contracts and second place on the broader LAB evaluation. [1] Although it is not yet deployable and lacks published weights or an API, Harvey indicated that the research will eventually integrate into its production enterprise platform. [1]

Harvey's shift toward post-training an open-weight base model like Kimi K3 signals a strategic push toward specialized, domain-specific AI ownership for enterprise law firms. By demonstrating substantial performance gains on internal benchmarks and establishing transferability to third-party evaluations like Mercor's APEX Agents, Harvey proves the viability of its asynchronous reinforcement learning methodology without relying on proprietary customer data. However, the current lack of deployable artifacts means immediate utility is limited to observing the methodology's potential. What the sources don't address: How the inference costs of this specialized model will compare to running general frontier models once it is fully integrated into Harvey's enterprise platform.