Thomson Reuters launches an in-house AI model built on open weights

Thomson Reuters announced Thomson, its first in-house large language model, built through continual learning and post-training on an existing open-weight foundation rather than pretraining from scratch. The company said this cost around $40 million in compute and staff time over two years; SiliconANGLE reported the final training run itself cost roughly $450,000 of that total. A technical report on the method, listing 26 authors, followed three days later under the title “Thomson: Continual Learning of Frontier Models for SovereignAI.”

Thomson draws on the company’s Westlaw, Practical Law, Checkpoint and Reuters content, with hundreds of subject-matter experts involved in training and evaluation, and will first be deployed inside Tabular Analysis, a document-review feature of its CoCounsel legal assistant. Thomson Reuters described the model as performing on par with recent frontier models across a range of tasks, a claim SiliconANGLE noted “has not yet received extensive independent validation.” A smaller open variant, Thomson-1.0-Small, was published on Hugging Face under a noncommercial licence; its model card states it was built through continual learning on top of Alibaba’s open-weight Qwen3.6-35B-A3B.

For years, the AI industry has treated scale as the answer: bigger models, more compute, more money. Thomson shows there is another path. Start with a strong foundation, specialize it deeply for the work that matters, and you can build intelligence that is highly capable, far more efficient and entirely under your control.

— Joel Hron, Thomson Reuters chief technology officer

The “SovereignAI” argument — that continual learning lets organisations outside the frontier labs reach competitive capability while keeping deployment and data under their own control — sits against a specific history for Thomson Reuters: the company won the first US ruling to reject an AI company’s fair-use defence for training data, against a legal-search rival that had trained on its Westlaw headnotes. Its own model now follows a pattern also seen days earlier at Harvey, which post-trained a legal-AI product on Moonshot’s open-weight Kimi K3: professional-services firms building competitive models by fine-tuning openly published weights rather than training from zero.