Thomson Reuters announced its first proprietary large language model, Thomson, on August 24, 2026. The model is not positioned as a general-purpose challenger in every field. Instead, it brings the company’s legal, tax, regulatory and news content, along with editorial and subject-matter expertise, into the training and evaluation process.
Where will Thomson be used first?
Thomson’s first product integration is Tabular Analysis in CoCounsel Legal, a feature that helps lawyers work through large volumes of documents and tables. SiliconANGLE reports that Thomson will be the default model for this feature, while CoCounsel will remain multi-model and administrators will be able to choose other models when needed.
The arrangement clarifies Thomson Reuters’ strategy: not every task goes to one model. A specialized model is used where the company has a strong content and tool advantage, while external frontier models remain part of the stack for other work. For enterprises, model selection can be designed alongside workflows and data permissions.
How was it built?
Thomson Reuters says Thomson started from an open-source foundation and then added content from Westlaw, Practical Law, Checkpoint and Reuters through mid-training and post-training. Hundreds of subject-matter experts helped define training objectives, create examples and judge blind evaluations. The model was also trained to work with tools such as Westlaw and Practical Law.
The company says the program involved about $40 million in investment, while the final training run cost about $450,000. Those figures describe Thomson Reuters’ own approach and claims; they do not mean every company can reproduce the result on the same budget. The important assets are not only model weights, but decades of professional content, editorial processes, tools and evaluators.
What evidence is available so far?
Thomson Reuters says early evaluations put Thomson close to leading frontier models on several tasks, with an advantage in legal completeness and citation quality when connected to its own content. The company also says less than 10% of its total information base has been used so far.
These results should still be separated from independent validation. The model was shared with legal and AI academics before launch and is expected to remain available for outside testing. A smaller open-weight version and a portal for developers to request API keys are planned next steps, not features that became broadly available with this launch.
Why should other AI teams care?
Thomson shifts the enterprise-model question from “can we build a larger general model?” to “can we turn our content, tools and professional judgment into useful work capability?” For finance, accounting, healthcare, manufacturing and legal teams, the real difference may be whether a model uses internal data correctly, respects permissions, provides traceable citations and hands control back to people when it is wrong. The name of a model is not the moat; productizing governance and workflow together is.
For official details, see Thomson Reuters’ official announcement、the Thomson model page，以及 SiliconANGLE’s product report。
