Thomson Reuters launches proprietary AI model

Thomson Reuters has launched Thomson, its first proprietary large language model, and plans to use it first in CoCounsel Legal.

Thomson Reuters announces in an official press release that it trained the model in house, using an open source base model and about $40 million in investment for talent and computing resources. The company says it owns and controls the resulting model and can run it at a lower cost than comparable frontier models.

Thomson was further trained using proprietary material from products including Westlaw, Practical Law, Checkpoint and Reuters. The company also involved hundreds of subject matter experts in defining training goals and evaluating the results.

Thomson Reuters says the model improves on its underlying foundation model in following detailed instructions and working with dense, specialised material. It argues that access to professional content alone is not enough for expert performance. In its view, specialised training and expert feedback are also needed.

The company has trained Thomson on less than 10 percent of its content so far. It plans to expand the model’s specialisation rather than simply add more data.

Its first use will be Tabular Analysis in CoCounsel Legal, a feature for reviewing structured information across large sets of documents. Thomson Reuters says CoCounsel will remain a multi model product, using Thomson for tasks where it performs best and other models for other functions.

The company is also making a smaller open weight version available through Hugging Face for academic and non commercial use. External legal and AI researchers are testing the model, though the company’s performance claims remain based on early evaluations.

Thomson Reuters also frames the launch as a move toward AI sovereignty. It says direct control over training, deployment and customer data handling matters for professionals working under legal, financial and other duties of care.