HARVEY BUILDS ITS OWN LEGAL AI ON A CHINESE OPEN-WEIGHT MODEL
Harvey built its first specialized legal model by post-training China’s Kimi K3 — showing how AI application companies can increasingly take powerful open-weight models, customize them for an industry and reduce their dependence on the frontier AI companies underneath them.
WHAT’S HAPPENING
Legal AI company Harvey has introduced Harvey Tenet, its first model post-trained specifically for legal work.
Instead of building a foundation model from scratch, Harvey started with Kimi K3, an open-weight model developed by Chinese AI company Moonshot AI, and worked with Fireworks AI to post-train it for complex legal tasks.
Harvey says Tenet produced significant improvements over the underlying Kimi model on its legal benchmarks, including long-horizon tasks requiring AI to work through large amounts of legal material.
WHY IT MATTERS
Harvey previously built its platform around access to general-purpose models from companies including OpenAI, Anthropic and Google.
Tenet demonstrates another path.
An application company can take an increasingly capable open-weight foundation model and build its own specialized intelligence on top of it instead of relying exclusively on closed models controlled by outside providers.
That could shift leverage throughout the AI industry.
The company closest to the customer may increasingly be able to own more of the intelligence itself.
WHO BENEFITS
Harvey potentially gains greater control over model behavior, customization and inference costs while developing AI specifically around legal workflows rather than depending entirely on general-purpose models.
Law firms could also eventually gain models trained more specifically around their own work, processes and institutional knowledge.
WHO LOSES
Frontier AI companies could face pressure if successful application companies determine they no longer need to route every specialized task through proprietary models.
Legal-AI competitors may also face higher barriers if Harvey can combine its customer relationships, proprietary legal workflows and specialized models into one system.
WHAT HAPPENS NEXT
Harvey says Tenet is part of a broader effort to allow legal organizations to develop specialized models around their own work.
The bigger question goes far beyond law.
If powerful open-weight models become good enough for companies to post-train themselves, more AI businesses may decide that the model they rent today can become the model they customize — and eventually control — tomorrow.
