Harvey Built Its Own Legal AI Model — and Stopped Outsourcing to OpenAI and Anthropic

Harvey, a legal AI company, launched its first in-house model called Tenet on August 18, 2026. The model was built specifically for legal reasoning. It wasn’t a general-purpose chatbot adapted for law. Harvey introduced Tenet as part of Harvey II, a broader product update tied to improved memory and legal workflows.

Before Tenet, Harvey relied on frontier models from OpenAI, Anthropic, Google, and Mistral. The company had used a multi-model setup to power its platform. Earlier reporting showed Harvey shifting away from a single-model reliance on OpenAI in 2025. It added Anthropic’s models that same year. But even with multiple providers, Harvey was still outsourcing its most important legal tasks to outside companies.

Tenet changes that. The model was reportedly built on top of Kimi K3, an open-weight model from Moonshot AI. Kimi K3 was released in July 2026 and contains 2.8 trillion parameters. Harvey then post-trained the model end-to-end for legal work. Reports indicated that Fireworks AI helped with training focused on long-horizon legal tasks. These include contract review, diligence, evidence sweeps, and litigation work that can take hours to complete.

Harvey claimed Tenet reached frontier-level results on major legal benchmarks. The company said it matched the performance of strong general models but at open-source cost. Product descriptions emphasized agentic legal work, including tool use and multi-step workflows across complex matters. The model was positioned as a legal-specialized step forward, not a broad foundation model. Harvey II also introduced agents that can process extensive document sets overnight, with costs recorded against the respective matter.

Harvey’s relationship with OpenAI goes back years. OpenAI once said Harvey partnered with it to build a custom-trained case law model. OpenAI remained part of Harvey’s platform stack even after the company diversified. But Tenet represents a shift toward self-owned capability. Harvey no longer needs to route every difficult legal task through OpenAI or Anthropic. The development process involved creating mock disputes and case files to capture how lawyers think and reason, with direct input from practicing attorneys.

The company framed Tenet as a way to cut costs. Running expensive legal tasks through outside models added up. With its own model, Harvey can control cost, performance, and specialization. This mirrors a broader industry trend where human oversight ensures that AI-driven decisions remain trustworthy and free from the biases that can emerge in automated systems. The move reflects a broader strategy centered on legal-specific workflows rather than dependence on generic large language models.