Thomson Reuters has launched Thomson, its first proprietary large language model developed in-house. The company says it began with a strong open-source foundation and then specialized the model using its professional content, tools, and subject-matter expertise. The launch pushes Thomson Reuters beyond supplying data to third-party AI systems and into owning a model it can operate and adapt itself.

The company reports spending $40 million across talent and compute to train Thomson. It says the model is fully owned and controlled by Thomson Reuters and can run at a fraction of the inference cost associated with typical frontier models. Those economics could be important in legal, tax, accounting, and compliance work, where organizations may process large volumes of sensitive material and need predictable operating costs.

What Thomson Reuters says makes the model different

  • Training draws on selected content and expertise connected to Westlaw, Practical Law, Checkpoint, and Reuters.
  • Subject-matter experts helped shape training objectives and evaluations for professional tasks.
  • The first deployment is planned for Tabular Analysis in CoCounsel Legal, while CoCounsel remains a multi-model product.
  • A smaller open-weight version is being offered on Hugging Face for academic and non-commercial evaluation.

This is not simply a company replacing every external model with its own. Thomson Reuters says CoCounsel will continue to use multiple models and apply Thomson where it offers the clearest advantage. That approach is sensible because legal research, document review, drafting, and general reasoning do not necessarily reward the same model. Routing each job to the right system can matter as much as having one impressive benchmark score.

The evidence still needs independent testing

Nearly all of the launch evidence currently comes from Thomson Reuters itself. The company says its early evaluations place Thomson alongside leading frontier models on a range of tasks and reports gains in instruction following and work with dense, domain-specific material. It has also begun giving legal and AI academics access for external evaluation. Until fuller methods, datasets, and reproducible results are available, the strongest performance claims should be treated as vendor-reported rather than settled comparisons.

The data story also deserves careful attention. Thomson Reuters says less than ten percent of its proprietary content has been used in training so far, but the launch announcement does not provide a complete public breakdown of the base model, dataset composition, model size, or every evaluation condition. Professional customers will also need clear answers about deployment location, confidentiality controls, audit trails, and how customer information is kept out of training unless consent is given.

The broader signal is that valuable enterprise AI may become more specialized. Publishers, research providers, and software companies with trusted data can use open foundations as a starting point, then build models around the language and constraints of their industries. The durable advantage will not come from owning a model alone. It will come from proving that specialized systems are more accurate, easier to verify, and cheaper to run in the workflows customers actually depend on.