Thomson Reuters spent $40M on a model only winning with its own data

thomson reuters built 'thomson', a proprietary legal model based on alibaba's qwen3.5-397b, investing $40m and two years to reduce dependence on api providers. internal benchmarks reveal the model is competitive but generally trails frontier models like gpt-5.4 unless it accesses the company's exclusive westlaw and reuters content, where it achieves a narrow lead. the firm prioritized owning the stack to control costs and integration, though the model's standalone performance lags significantly in coding and reasoning. future gains depend on increasing training data beyond the current 10 percent usage and updating to stronger base architectures.