Thomson Reuters has officially launched Thomson, its proprietary large language model trained on decades of its own legal content. Developed for a fraction of the cost of frontier models like ChatGPT and Claude, Thomson is set to power the company's CoCounsel Legal product.
TL;DR
- Thomson Reuters launches Thomson, an LLM trained on its proprietary legal content, developed for $40M.
- Thomson will first be deployed in CoCounsel Legal's Tabular Analysis functionality.
- The model is designed to be highly efficient and specialized for legal work, with plans to expand across TR's portfolio.
What happened
Thomson Reuters has launched Thomson, its proprietary large language model (LLM), trained on its own legal content from Westlaw, Practical Law, and other sources. The model was developed over two years with an investment of $40 million, significantly less than the billions spent on models like ChatGPT and Claude. The final training run for the version launching today cost just $450,000, according to Joel Hron, TR's chief technology officer.
Thomson will first be deployed in the upcoming release of CoCounsel Legal, driving its Tabular Analysis functionality. It will be the default model for this feature but can be switched to other models through admin settings. TR plans to extend Thomson models across its portfolio of legal and tax products over time.
Why it matters
Thomson's development highlights a cost-effective approach to building specialized AI models. By focusing on a strong foundation and deep specialization, TR demonstrates that high capability and efficiency can be achieved without the massive scale and cost of general-purpose models.
For developers and startups, Thomson's launch underscores the potential of domain-specific models. It also signals TR's commitment to integrating its proprietary content and expertise into AI solutions, which could influence the competitive landscape in legal tech and professional services.
Investors should note TR's strategic move to own and control its AI infrastructure. This could enhance its market position and drive future growth, particularly as the company explores licensing Thomson to large law firms and corporations.
Key facts
- Thomson was developed for $40 million over two years, with the final training run costing $450,000.
- The model is trained on Thomson Reuters' proprietary legal content from Westlaw, Practical Law, and other sources.
- Thomson will first be deployed in CoCounsel Legal's Tabular Analysis functionality.
- TR plans to extend Thomson models across its portfolio of legal and tax products.
- The model is designed to be highly efficient and specialized for legal work.
- TR has begun early conversations with large law firms and corporations about direct access to the model.
- Thomson is built to be repeatable, allowing TR to swap out its open-source starting point as better models emerge.
- TR will publish a technical report with comprehensive evaluations and make a small version of Thomson available as an open-weight model on Hugging Face.
Context
Thomson Reuters' launch of Thomson comes at a time when many companies are investing heavily in developing their own AI models. The focus on specialized, domain-specific models like Thomson could represent a shift in the AI industry, moving away from the one-size-fits-all approach of general-purpose models.
The legal tech industry is increasingly adopting AI to enhance efficiency and accuracy. Thomson's integration into CoCounsel Legal and potential expansion across TR's portfolio could set a new standard for AI-powered legal solutions.
TR's strategic move to own and control its AI infrastructure highlights the growing importance of data sovereignty and the need for companies to have control over their AI models and the data they are trained on.
