Thomson Reuters launches Thomson-1 model built on Alibaba Qwen open weights
On August 25, Thomson Reuters launched Thomson-1 built on Alibaba Qwen3.5; used for table analysis and document review, gradually reducing reliance on Claude.
On August 25, Thomson Reuters formally launched Thomson-1, a self-developed large language model built on Alibaba's Qwen3.5 open-source family. The model was trained on the firm's proprietary Westlaw and Reuters archives with pretraining, post-training, and reinforcement learning, and is designed to support the company's core legal and financial data business under tighter control.
Thomson-1 is being deployed first for table analysis and document review, gradually absorbing tasks previously handled by Claude. The model scores 0.823 on Stanford LegalBench and slightly below Opus 4.8 on the Harvey Legal Agent Benchmark. The company cited long-term cost and limited customisation as reasons for not relying on US closed-weight models.
Other enterprises have cited similar economics: Airbnb CEO Brian Chesky has publicly said his company leans heavily on Qwen, while Pinterest CEO Bill Ready has stated Qwen costs less than 8% of comparable closed-weight models. As of August, Qwen has open-sourced over 460 models with cumulative downloads exceeding 3 billion and more than 300,000 derivative models.