Harvey launches Tenet, its first proprietary AI model for law, built on China's Kimi K3

Summary

Harvey, an AI legal platform serving Magic Circle and AmLaw 100 firms, launched Tenet, its first proprietary legal LLM. The model was fine-tuned on the open weights of Kimi K3, a China-based model by Moonshot AI. This move immediately sparked debate over whether a China-origin base model poses structural risks to clients' confidential data, even if there is no technical backdoor.

Why it matters

Since Harvey processes attorney-client privileged information for the world's leading law firms, the origin of the model weights is not just a technical issue but a matter of compliance and reputation. This choice suggests that, in the context of the US-China AI decoupling, at least one major legal tech player prioritizes benchmark performance and cost over geopolitical alignment. With regulatory frameworks for LLM supply chain provenance in the legal sector yet to be written, firms are forced to make risk decisions in a vacuum.

Highlights

• Tenet is Harvey's first in-house model. It was fine-tuned on legal domain data (contracts, case law, regulations) using Kimi K3 open weights to reduce reliance on third-party APIs.
• Harvey serves Magic Circle and AmLaw 100 clients. While the article argues there is no technical data access vector for Moonshot, the China-origin base model raises compliance questions.
• Competitors include Thomson Reuters CoCounsel (GPT-4o based), LexisNexis Lexis+, Clio, Ironclad, and Luminance. Most of these rely on general-purpose models rather than legally fine-tuned weights.

Critical take

The article relies on a car engine analogy and the argument that 'open weights are just math,' ignoring the fact that corporate procurement and compliance teams are increasingly demanding full supply chain provenance documentation. Although the article acknowledges that such documentation is 'yet to be written,' it frames Harvey's choice as a bold geopolitical statement. The tension created by an OpenAI-backed firm building its flagship model on a China-origin base goes beyond a simple 'efficiency' issue; the article does not adequately confront this tension.