Thomson Reuters builds $40M legal AI model

Thomson Reuters has developed its first in-house artificial intelligence model, named Thomson, to handle specialized legal tasks. Built on Alibaba's open-source Qwen3.5-397B foundation, the model cost approximately $40 million in staff and computing power over more than two years, with the final training run alone costing $450,000. To ensure safety and political neutrality, the company worked with Imperial College to create an intermediate version called Snowdon before training the final model on proprietary data from Westlaw, Practical Law, Checkpoint, and Reuters.

The model's performance varies depending on its access to proprietary data. On the Stanford LegalBench, Thomson scored 0.823, trailing behind Gemini 3.1 Pro and GPT-5.5. It also placed just behind Opus 4.8 on the Harvey Legal Agent Benchmark. In factual accuracy tests on the company's Deep Research benchmark with web access alone, Thomson scored 0.53 compared to GPT 5.4's 0.65. However, when granted access to Thomson Reuters' exclusive internal content, Thomson edged out GPT 5.4 with a score of 0.83 to 0.82. The model leads in instruction following and PrBench Legal, though its reasoning and coding capabilities fall off sharply.

The company chose to build its own model rather than fine-tune frontier models from OpenAI or Anthropic to avoid high inference costs, vendor lock-in, and the degradation of general capabilities associated with standard fine-tuning. By training the model inside its own tool environments using reinforcement learning, Thomson Reuters retains full control over its data. This strategy allows the company to build long-term equity, treating AI development as an investment that compounds over time as expert reviews are fed back into the system.

For legal practitioners, Thomson will initially power the Tabular Analysis feature in CoCounsel Legal, handling high-volume subtasks like citation checking and document review. While the platform remains multi-model, administrators can choose to toggle Thomson on or off. Thomson Reuters is also releasing a smaller version of the model on Hugging Face under a non-commercial license, alongside a technical report, to foster open-source development.