Thomson Reuters AI: New Legal LLM "Thomson" Launched

Thomson Reuters has launched its own large language model called "Thomson," specifically designed for legal work. This model supports tasks related to legal advice by combining Thomson Reuters' legal content with external large language models. Thomson will first be integrated into CoCounsel's "table-based analysis," a feature for reviewing large documents. CoCounsel will maintain a multi-model approach, using Thomson for its specialized legal strengths and external models for other tasks. For table-based analysis, Thomson is the default, but administrators can select other models. Thomson Reuters invested approximately $40 million in staff and computing resources over two years for this project. However, the final training cost was reduced to about $450,000. This was achieved by building upon an open-weights model rather than starting from scratch, which lowered training and inference costs. Internal tests showed Thomson performed competitively against major models even when only accessing the web. Thomson Reuters also states it does not use customer data to train the model and believes owning its own model offers greater control over deployment and future development. This development could mean more specialized and cost-effective AI tools for legal professionals.