Thomson Reuters has launched Thomson, a proprietary large language model built for legal, tax and compliance work, the company announced on August 24. The model draws on decades of content from Westlaw, Practical Law, Checkpoint and Reuters news, with hundreds of subject-matter experts involved throughout training, the company said.

Built cheap, on a borrowed foundation

Rather than training an LLM from scratch, Thomson Reuters started from an open-weight base model and layered further training and domain specialization on top — an approach chief technology officer Joel Hron said keeps the result “highly capable, far more efficient and entirely under your control.” The company put its total spending, including staff and computing over two years, at about $40 million; trade press reported the final training run itself cost roughly $450,000 once those efficiencies kicked in.

Thomson Reuters’ own announcement described only “a strong open-source foundation,” but the model’s Hugging Face listing and Hron’s comments to other outlets identify the base as Alibaba’s Qwen3.6, an open-weight model from China, which Thomson Reuters and Imperial College London researchers adapted over several months for safety and neutrality. The choice underscores how even large Western enterprises are turning to Chinese open-weight models to cut costs — a trend that has already drawn scrutiny from US lawmakers over data and security risks.

Document review first, sovereign AI later

Thomson’s first production use is inside Tabular Analysis, a document-review feature in CoCounsel Legal, the company’s legal AI assistant already used by law firms and corporate legal departments. Thomson Reuters plans to extend the model across its legal and tax product lines, including future “sovereign AI” options that keep data within a customer’s own infrastructure — part of a broader push by enterprises to control their own AI stack rather than depend solely on outside frontier labs.

A smaller, open-weight version, Thomson-1.0-Small, is available on Hugging Face under a noncommercial academic license; wider enterprise access is rolling out through an upcoming CoCounsel Legal release. Thomson Reuters said the model has drawn on less than 10% of its content archive so far and that customer data is not used in training.

Read also