AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

OpenAI-Backed Legal AI Firm Harvey Builds Model on Chinese Kimi K3

OpenAI-backed legal AI company Harvey launches Tenet, a model based on Chinese AI Kimi K3, highlighting US AI startups' growing reliance on Chinese open-weight models.

AI News Feed · 2026-08-26 00:44

Harvey, founded just four years ago, provides AI tools for law firms and corporate legal departments, helping lawyers with legal research, contract review, and document processing. The company is valued at $11 billion and serves more than 1,300 institutions and 100,000 lawyers. Tenet is designed for long-horizon legal tasks that require continuous tool use. Initial results show that Tenet completes nearly twice as many test tasks as the base model on Harvey's legal intelligence benchmark LAB.

The decision to train a dedicated model stems from the limitations of general-purpose models. Lawyers face messy client documents and need models that can decide where to start, retrieve and read material sequentially, identify risks, cite sources, and produce deliverable work. These tacit skills are hard to encode in prompts or external workflows, making fine-tuning necessary.

Harvey is not alone. Cursor, the popular AI coding tool, was found to have built its Composer 2 model on Kimi K2.5, and its Composer 2.5 still uses a Kimi base. Cosine, a UK-based AI company, used Kimi K2.6 to train Lumen Outpost. Devin, once a star AI agent product, publicly acknowledged using Kimi K2.7 for its SWE-1.7 model. Within six months, different versions of Kimi have been picked up by overseas companies in sequence.

Thomson Reuters also released its first self-developed model, Thomson-1, using a Chinese open-source model. Its CTO Joel Hron described the choice as renting versus buying – calling external APIs rent, while owning a model lets companies turn their industry data and intellectual property into assets. Thinking Machines, founded by OpenAI's former CTO Mira Murati, used DeepSeek's V3 for architecture and Kimi K2.5-generated synthetic data to kickstart supervised fine-tuning.

These companies still invest in data, compute, and domain expertise to turn a general model into a usable product, but they no longer insist on claiming complete originality. The open-weight movement from China offers a third path beyond costly from-scratch training and recurring API fees, allowing firms to customize models for specific customers. As a result, Chinese models are increasingly serving as the raw materials and upstream industrial products for American AI applications, often unbeknownst to end users.