Thomson Reuters built its own model on a Chinese open base

Thomson Reuters launched its own large-language model, Thomson, after spending $40 million on talent and compute, building it on an adapted version of Alibaba’s open-source Qwen 3.5 model called Snowdon. The model, trained on under 10% of Reuters’ legal and tax content, powers the CoCounsel Legal AI assistant and will be expanded across the company’s legal-tax portfolio.