Thomson Reuters has launched an in-house large language model, "Thomson", specialised for legal work. SiliconANGLE reported on Aug. 24 that Thomson supports tasks related to legal advice by using Thomson Reuters legal content together with LLMs from external providers.

Thomson will first be applied to the legal AI assistant CoCounsel’s "table-based analysis", a large-scale document review feature. CoCounsel will keep a multi-model structure. Thomson Reuters plans to use Thomson for tasks where it has strengths in legal specialisation and leading external models for other tasks. In "table-based analysis", Thomson is set as the default model, and administrators can choose other models.

Thomson Reuters invested about $40 million in staff and computing resources in the project over 2 years. But it lowered the final training cost to about $450,000. It explained that by not building a base model from scratch and instead adding its content, training techniques and expertise to an open-weights model, it reduced training and inference costs compared with general-purpose leading models.

Thomson Reuters said internal tests showed Thomson delivered overall competitive performance against major models when it had access only to the web.

Thomson Reuters said it does not use customer data to train the model. It also explained that owning its own model allows it to more directly control deployment, governance and future development.