Thomson Reuters launches in-house LLM ‘Thomson’

Thomson Reuters has developed an in-house large language model called Thomson. The model was trained on decades of the company’s legal and tax content after the firm acquired Safe Sign Technologies in August 2024 and invested roughly $40 million in training runs.

Executives say Thomson was built to support specialist professional markets where high accuracy is required. Legal and tax customers have demanded near-perfect answers for research, drafting and document review rather than general-purpose, probabilistic outputs. Kirsty Roth, chief operating officer, said the technology giants will remain dominant for broad use cases but that specialized verticals need focused models.

Thomson was developed on an open-source foundation and so far has been trained on less than 10% of Thomson Reuters’ proprietary content. The company describes training as the main cost driver. It estimates an individual training run can cost about $400,000 and said total training investment has reached the tens of millions.

The model’s first customer-facing use is inside Tabular Analysis in CoCounsel Legal, an AI assistant for large-scale document analysis such as due diligence and for producing concise summaries for attorneys. Internally, engineers are using the model to help write code and to speed product release cycles; Roth noted AI now contributes to a large portion of the firm’s codebase.

Thomson Reuters says the smaller, domain-focused model offers lower latency and lower inference costs than larger public models, and gives the company full control over its model and the data it uses. The firm cited improved operational control and fewer third-party dependencies as reasons for running an in-house model.

The company has not committed to making Thomson available for external sale, though it acknowledged interest from outside firms. Thomson Reuters plans to continue development, adding more content and features in future versions. The effort is part of a wider pattern of established companies building sector-specific AI tools instead of relying only on large public models.