24 August 2026

Thomson Reuters builds proprietary legal AI model after $40M investment

• Thomson Reuters launched Thomson, an in-house language model built on Alibaba's open-source Qwen and trained partly on its decades of legal content from Westlaw and other platforms.
• The company spent roughly $40 million over two years on staff and computing power, though the final training run alone cost $450,000, according to Thomson Reuters.
• Thomson performs competitively on legal benchmarks when accessing the company's proprietary tools and content, but trails GPT and Gemini models on general reasoning and coding tasks.
• The company chose to build rather than fine-tune external models because training on its own workflows improved performance more than general capability, and it retains long-term ownership and data control.
• Thomson Reuters argues this approach works only for companies with exclusive data, domain experts on staff, and measurable quality metrics, but shows open-source models can compete with frontier labs in specialized domains.

Reported by The Decoder