Thomson Reuters has built an in-house language model called Thomson for legal and professional work, spending about $40 million on staff and computing over more than two years. The model starts from Alibaba's open Qwen family, then adds company content, domain-expert training and practice with internal tools.
Its strongest reported results depend on those private resources. On an internal research benchmark using only the web, Thomson scored 0.53 for factual accuracy against 0.65 for GPT-5.4. With Thomson Reuters content, it narrowly led 0.83 to 0.82. The model also performed well on some legal and instruction-following tests but lagged on reasoning and especially coding. Differences in test-time reasoning make some comparisons uneven.
The first planned use is document review, and a smaller version will be available under a non-commercial license. The project suggests that owning a model can provide cost and independence benefits, but its clearest asset may be the surrounding model factory, proprietary corpus and expert workflow—not a universally superior base model.