Thomson Reuters released its own large language model this week, called Thomson, adapted from Alibaba's open-weight Qwen and aimed at cutting what the company pays outside AI providers. Thomson is trained for journalism, law, and tax work rather than coding, and a second model screens the base technology against Thomson Reuters' own standards on safety, ethics, and political neutrality, though executives said they could swap in a different foundation model if a better one comes along. CEO Steve Hasker told investors this month that training cost $40M, while CTO Joel Hron put the final training run at roughly $450,000, per The Logic. Developers can test the model on Hugging Face, and CoCounsel, the company's legal chatbot, starts routing some advanced features to Thomson.
