Thomson Reuters launches Thomson model for professional work
Thomson Reuters launched Thomson, its first proprietary large language model, on August 24. The company developed the model in-house for legal, tax and other professional work, starting from an open-weight foundation and investing $40 million in talent and computing. Thomson Reuters says it owns and controls the resulting model rather than depending entirely on an outside frontier provider.
A companion technical account identifies the current base as Snowdon, an open-weight model developed by Imperial College London's FAIR Lab, which Thomson Reuters helped establish. The company then used mid-training and post-training with selected material from Westlaw, Practical Law, Checkpoint and Reuters, plus evaluation and preference work by subject-matter experts. Less than 10 percent of its content has been used so far.
The first production deployment is in Tabular Analysis inside CoCounsel Legal, where the model will support high-volume, structured document review. Thomson Reuters says it will enter production this month and later expand across its legal and tax portfolio. CoCounsel remains a multi-model product: Thomson will handle tasks where the company sees a specific advantage, while other models remain available elsewhere. A smaller Thomson version is also being released with open weights on Hugging Face for academic and non-commercial evaluation.
The launch is notable because it tests a different route to competitive AI. Instead of training a general-purpose model from scratch at the scale of the largest labs, Thomson Reuters specialized an open foundation with proprietary data and professional expertise. That could give organisations with deep archives another option between renting closed models and accepting the limitations of an unchanged open model.
Performance claims still require caution. Thomson Reuters says internal evaluations place Thomson alongside leading frontier systems and report stronger citation factuality in legal research. Those results were designed or published by the company and are not independent proof of reliability across customer work. External academics have begun testing the model, while a full technical report is still forthcoming.
For legal and tax teams, the immediate benefit could be greater control over model behaviour, data location and costs. The responsibility does not move to the model, however. Professionals still need to verify sources, review outputs and remain accountable for advice. The release shows that competition in enterprise AI is increasingly about specialised training, trusted data and deployment control, not only model size.