Thomson Reuters is launching “Thomson,” its first in-house language model, according to The Decoder. The model is built on Alibaba’s open Qwen family, with the report citing Qwen3.5-397B as the latest foundation used by the company. The investment is materially larger than the training-run number that has circulated around the project. The Decoder reports, citing the company, that Thomson Reuters spent about $40 million on staff and compute over more than two years. The roughly $450,000 figure applies only to the final training run for the current version. That accounting still understates the main asset behind the system: Thomson Reuters’ proprietary corpus. The Decoder says the model can draw on decades of material from Westlaw, Practical Law, Checkpoint and Reuters, as well as work from hundreds of domain experts. So far, less than 10% of the available content has gone into training. The development path was not simply a fine-tune on legal text. According to The Decoder, Thomson Reuters worked with Imperial College to retrain the Qwen base model for safety, ethics and political neutrality, producing an intermediate model called “Snowdon.” The company then added pre-training on its own content, post-training with domain experts and agentic reinforcement learning inside its own software environments. The benchmark picture is mixed. The Decoder reports that Thomson Reuters’ public framing places Thomson among top models, but the company’s numbers show a more specific result: on Stanford LegalBench, Thomson scored 0.823 and trailed Gemini 3.1 Pro and GPT-5.5. On the Harvey Legal Agent Benchmark, it placed just behind Opus 4.8. The report says it led on instruction following and PrBench Legal, while falling off on reasoning and especially coding. The clearest advantage appears when the model can use Thomson Reuters’ own content. In an in-house Deep Research benchmark, The Decoder says Thomson scored 0.53 on factual accuracy with web access alone, below GPT 5.4 at 0.65. With access to company content, Thomson edged GPT 5.4 by 0.83 to 0.82. Evaluation lead Andrew Bean told The Decoder that web-only performance was within the range of other models but “certainly not the leader yet.” That distinction matters for how to read the strategy. The report suggests Thomson Reuters is not claiming a general-purpose frontier model that beats the largest outside systems across the board. It is building a domain system whose value comes from tight access to proprietary material, workflows and evaluation environments that outside providers cannot automatically replicate. Executives also described the project as a capability-building exercise. CTO Joel Hron told The Decoder that Thomson Reuters has changed the open-source starting point “close to a half dozen times,” while research chief Jonathan Schwartz said the more important outcome is the company’s “model factory.” Schwartz also argued that standard fine-tuning can degrade general capability and still leaves customers tied to a provider’s inference costs and roadmap. For now, the evidence supports a narrower conclusion than the headline number alone implies: Thomson Reuters has put real money and proprietary data behind an in-house AI stack, but the model’s best showing depends on the same proprietary content that defines the company’s moat. Who benefits: Thomson Reuters benefits if its content and tool access produce better legal, tax and professional workflows than a generic model interface. Users of products such as Westlaw or Practical Law may benefit if those gains translate into more accurate in-product AI assistance. Who's exposed: Outside model providers are exposed where large content owners can justify building and operating their own models. Thomson Reuters is exposed if its in-house model remains dependent on proprietary context for narrow benchmark wins while trailing stronger models in general reasoning or coding.