Thomson Reuters Develops Proprietary AI Model Based on Alibaba's Qwen

Thomson Reuters has reportedly developed its own large language model named "Thomson" based on Alibaba's Qwen base model, costing approximately $40 million to build. The company chose to build its own model to avoid the high long-term costs and lack of customization associated with US closed-source alternatives. This highlights a growing trend of major enterprises opting for open-source foundations to build proprietary, domain-specific models for better cost control and data privacy. However, the report's credibility is questionable due to references to non-existent model versions like "Qwen3.5-397B" and "GPT-5.5". The model achieved a score of 0.823 on Stanford's LegalBench, though this evaluation utilized inference-time scaling, which was not applied to the competing models. Additionally, the reported $40 million development cost contrasts with a widely circulated $450,000 figure, which only represented the final training run.