Thomson Reuters has launched Thomson, its first proprietary language model — a signal that specialized companies can build competitive AI systems without spending billions on foundation model research. The model was built from an open-source base and specialized using proprietary data from Westlaw, Practical Law, Checkpoint, and Reuters.
The company invested approximately $40 million in the project, a fraction of what major AI labs spend on foundation models. Thomson Reuters trained the model using its legal, tax, and professional expertise, and integrated it with the workflows and tools in which it will ultimately be used. The model is already prepared for integration into CoCounsel Legal, the company's AI-powered legal research platform.
Early evaluations place Thomson at the level of the latest frontier models on a range of tasks, according to the company. What sets it apart is not raw capability but specialization: the model was trained on domain-specific data that general-purpose models cannot access, giving it an edge in legal and professional research contexts.
For the broader AI industry, the launch demonstrates a model that could challenge the assumption that only well-funded AI labs can produce competitive language models. Thomson Reuters owns specialized data, domain experts, and the end-user tools in which the model will be deployed — a combination that allows it to start with an existing foundation model and transform it into a highly effective system for a specific industry.
The approach is also a concrete example of what some analysts call corporate AI sovereignty: the company controls the model, the training data, and the integration path. This contrasts with the approach of licensing a third-party model, where the company has limited control over how the model evolves, how its data is used, and how it integrates with existing products.
