Thomson reuters develops custom AI model using alibaba's Qwen

Thomson Reuters, the parent company of Reuters, has unveiled its first large language model, named 'Thomson,' which is built on Alibaba's Qwen. The development of the model cost approximately $40 million, with the previously reported figure of $450,000 referring only to the final training phase. The model is based on the Qwen3.5-397B version of Alibaba's Qwen and was initially retrained by the company in collaboration with Imperial College London to meet safety, ethical, and political neutrality standards. Thomson Reuters then further trained the model using its own content and conducted reinforcement learning within its internal environment. The company claims the model ranks among the world's best AI models, achieving a score of 0.823 in the Stanford LegalBench test, which is slightly behind Gemini 3.1 Pro and GPT-5.5. In the Harvey Legal Agent Benchmark, Thomson scored just below Opus 4.8. However, the model's performance in the test was enhanced through inference-stage expansion, while GPT-5.5 was tested without this feature. In a deep search test, Thomson achieved a score of 0.53, compared to GPT 5.4's 0.65. Thomson Reuters stated that it opted for developing its own model rather than using closed-source models from companies like OpenAI or Anthropic due to higher long-term costs and limited customization options. The company emphasized that a smaller, self-developed model is more suitable for tasks such as document review.