Thomson Reuters has spent $40 million over two years to build its own specialized legal model, dubbed "Thomson," built on top of Alibaba’s open-weights Qwen architecture. Rather than writing open-ended checks for API tokens to OpenAI or Anthropic, the publishing giant chose to construct its own stack. The calculation is straightforward: lower long-term inference operational expenses, maintain absolute data sovereignty, and automate heavy document review workflows without shipping sensitive customer records to third-party endpoints.
To bridge the gap between base weights and enterprise legal reasoning, Thomson Reuters collaborated with Imperial College London to align an intermediate checkpoint called "Snowdon." That model was then fine-tuned on proprietary archives from Westlaw, Practical Law, Checkpoint, and Reuters alongside hundreds of domain experts. Notably, the team notes that less than 10 percent of its proprietary data moat has actually been utilized in training so far.
Yet the raw technical results expose the real bottleneck in modern enterprise AI: domain architecture is nothing without walled data. Evaluation lead Andrew Bean reported that on an internal Deep Research benchmark with standard web access, Thomson trailed significantly with a factual accuracy score of 0.53 compared to GPT-5.4’s 0.65. The custom model only pulled ahead—by a razor-thin margin of 0.83 to 0.82—when tethered directly to Thomson Reuters' proprietary databases. For enterprise leaders eyeing a custom model strategy, the lesson is stark: owning the weights will not save you unless you already own the definitive data moat.