The company invested $40 million to train the model, utilizing a foundation of open-source architecture refined through decades of proprietary data from Westlaw, Practical Law, and Reuters. Unlike general-purpose models that prioritize raw scale, Thomson is specifically engineered for domain-specific reasoning and complex instruction following. CTO Joel Hron noted that the strategy focuses on deep specialization rather than brute-force compute, enabling the firm to maintain full control over the model's behavior and data privacy.
In section Releases
Thomson Reuters Debuts Proprietary AI Model to Rival Frontier LLMs
Thomson Reuters has launched Thomson, a proprietary large language model developed in-house for a fraction of the cost associated with current industry leaders. By pivoting away from the industry's reliance on massive compute spending, the company aims to establish a new standard for sovereign, professional-grade artificial intelligence.
Early performance evaluations indicate the model is competitive with top-tier frontier alternatives, particularly in legal and tax-related tasks. The company has begun integrating the technology into CoCounsel Legal, specifically for tabular document analysis. To facilitate external validation, Thomson Reuters is releasing a smaller, open-weight version of the model on Hugging Face for academic and non-commercial research. This shift toward in-house development marks a strategic move to secure AI sovereignty, allowing the firm to offer professional tools that meet strict fiduciary-grade standards for accuracy and accountability.
Comments (0)
No comments yet. Be the first!