Harvey Unveils Tenet Model for Legal Agent Work

Legal technology company Harvey has announced the rollout of Harvey Tenet, marking the organization's very first post-trained model available as a research preview. The system relies on a Kimi K3 base that has undergone post-training through Fireworks, utilizing asynchronous reinforcement learning specifically tailored for complex, multi-step legal workflows. According to the organization, the training corpus incorporated a blend of synthetic data, publicly available legal documents, and human expert materials, while strictly avoiding any customer data during the process.

The technical foundation behind the model relies heavily on asynchronous reinforcement learning executed within sandboxed legal environments. These environments mirror the structure of Harvey's Legal Agent Benchmark, featuring detailed partner-style instructions, relevant client documents, and expert rubrics containing numerous atomic pass-fail criteria. During optimization, individual rollouts can extend past one thousand turns. The grading process employs an LLM-as-a-judge mechanism, while policy optimization is handled via GSPO utilizing a rank-64 LoRA applied across the entire K3 network architecture.

This development arrives during a period of rapid advancement in domain-specific artificial intelligence, where general-purpose foundational systems are increasingly adapted into specialized vertical agents. Law firms and legal tech providers have been searching for architectures capable of managing long-horizon reasoning, which requires maintaining context and executing multi-step workflows without drifting from professional standards. By building on top of open-weight foundations, the initiative aims to provide legal institutions with a viable path toward owning specialized models rather than relying entirely on closed ecosystem providers.

When evaluated against the base K3 model, Harvey Tenet demonstrates notable improvements on internal metrics. The system completes nearly twice as many held-out tasks on the Legal Agent Benchmark and improves performance on contract evaluations, securing state-of-the-art results in specific contract categories while achieving strong placements overall. Crucially, these performance gains transferred effectively to external evaluations such as Mercor's APEX Agents and Crosby's Redline Bench, neither of which were included in the training data, while preserving baseline legal reasoning capabilities on standard academic tests.

Despite these achievements, several limitations and constraints accompany the current release. At this stage, Harvey Tenet remains strictly a research preview, meaning the organization has not yet published model weights, comprehensive model cards, or accessible API endpoints. What has been shared is primarily the training recipe and methodology rather than an immediately deployable commercial artifact. Furthermore, much of the performance validation relies on specialized benchmarks, and independent verification across broader real-world legal practices will be necessary to confirm the robustness of long-horizon execution.

Looking ahead, the organization plans to transition this research framework into production systems integrated directly into Harvey's commercial offerings over time. Observers and legal technology developers will be watching for the eventual release of model weights or APIs, as well as broader third-party benchmarks that can independently verify how well asynchronous reinforcement learning scales for complex professional services. The success of this approach may influence how other vertical industries attempt to customize open-weight foundation models for specialized expert tasks.