Harvey Unveils Tenet, a Kimi K3 Legal Agent Model
Legal technology company Harvey has introduced Harvey Tenet, a large language model built by post-training a Kimi K3 base model for extended, multi-step legal work, according to a report from Marktechpost.
Per the report, Harvey partnered with AI infrastructure provider Fireworks to carry out the post-training process. The resulting model is designed to function as an agent capable of handling long-horizon legal tasks, meaning workflows that require sustained reasoning and multiple sequential steps rather than single-turn responses.
Marktechpost’s report frames Tenet as part of a broader trend of specialized firms adapting open base models for domain-specific agentic use cases, in this case tailoring a Kimi K3 foundation toward legal applications such as document review, research, and drafting tasks that unfold over many steps.
The move reflects growing interest among enterprise AI companies in building on strong open or semi-open base models rather than training foundation models from scratch, then investing in post-training techniques to specialize them for narrow, high-value professional domains like law. Fireworks’ involvement points to the increasing role third-party infrastructure providers play in helping companies fine-tune and deploy large models at scale without building their own training stacks.
Marktechpost’s coverage did not detail specific benchmark results, parameter counts, or a public release timeline for Harvey Tenet. The report centers on the model’s origin as a post-trained variant of Kimi K3 and its intended use for agentic legal workflows, positioning it within Harvey’s existing product line aimed at law firms and corporate legal departments.
The development underscores continued momentum in applying large language models to specialized professional services, with legal AI remaining one of the more active verticals for agent-style deployments that go beyond simple question answering into extended task execution.
Based on reporting by www.marktechpost.com.
