Thomson Reuters Launches Thomson, Its Own Legal AI Model Built on Alibaba's Qwen

Thomson Reuters has launched its first large language model, Thomson, built on Alibaba's Qwen architecture, with development costs around $40 million. The company first worked with Imperial College to adjust the base model for safety and ethics, then pre-trained and reinforced it with its own content.

In professional legal tests, Thomson performed well, scoring 0.823 on Stanford LegalBench, just behind Gemini 3.1 Pro and GPT-5.5, and slightly below Opus 4.8 on the Harvey Legal Agent Benchmark. However, Thomson used inference-time scaling during testing, while GPT-5.5 did not.

Thomson Reuters said it chose not to use closed models from OpenAI or Anthropic because external models have higher long-term costs and less flexibility. A smaller custom model is better suited for tasks like document review.