Thomson 1.0: Continual Learning on Open Weights Nears Frontier for a Fraction of Big-Lab Budgets
Thomson Reuters built a frontier-grade model for roughly $40M by continual-learning an open-weight base instead of training from scratch. The technical report, co-written with Imperial College's FAIR Lab, shows gains comparable to several generations of model updates at 0.3x the typical cost-per-test of rivals. Open-weight Thomson-1.0-Small (35b) ships as a template for "sovereign AI," with π-shaped gains that avoid the forgetting that usually hits narrow fine-tuning.