模型與企業 AI
Thomson Reuters Launches Thomson Legal Model: Continued Training from Snowdon Weights, Debuting in CoCounsel Document Review
Thomson is based on Imperial College London’s open-weight Snowdon model, with mid-training and post-training using proprietary data from Westlaw, Practical Law, and other sources, as well as feedback from legal experts. Official evaluations show better citation verifiability than general-purpose frontier models, but the full technical report and independent validation are still pending.
Thomson Reuters has officially launched Thomson, its proprietary large language model. Its first production deployment is in Tabular Analysis for CoCounsel Legal, where it supports structured review of large document sets. CoCounsel has not shifted to a single-model architecture: Thomson will handle tasks where it has a domain-specific advantage, while other tasks may still be routed to third-party frontier models. Administrators can also select alternative models.
This is not a foundation model pretrained from scratch. According to the company’s disclosures, the current version starts from Snowdon, an open-weight model developed by Imperial College London’s FAIR Lab. It underwent continued pre-training on selected, cleaned, deduplicated, and blended data from Westlaw, Practical Law, Checkpoint, and Reuters News, followed by post-training using complex scoring rubrics created by legal experts, preference data, and tool-use trajectories. So far, the Thomson Reuters content used represents less than 10% of the company’s total corpus.
The team made preventing “catastrophic forgetting” during domain fine-tuning a training objective, with particular attention to instruction-following capabilities, which tend to degrade early. In the company’s results, Thomson achieved an aggregate instruction-following score of 0.914. On a deep-research test comprising 53 internal legal-research questions, Thomson scored 0.83 for factuality after being connected to Westlaw and Practical Law, based on item-by-item checks of whether each claim was supported by its cited source. Two frontier systems allowed to search the public web freely scored 0.65 and 0.68.
The project involved approximately $40 million in investment over two years, covering talent and compute. Reports put the cost of the final training run at about $450,000. This approach has practical significance for companies that hold high-quality proprietary corpora: competitive differentiation may shift from parameter count toward data rights, expert-designed scoring rubrics, and reproducible domain evaluations.
However, the published tables combine industry benchmarks, internal long-context tests, and model-as-a-judge evaluations. The model’s size, training-token count, and the full specifications of the underlying Snowdon model have not been disclosed. A complete technical report has yet to be published, while a smaller open-weight version and an external API remain only planned. At this stage, the vendor’s evaluations should not be interpreted as proof that Thomson comprehensively outperforms Claude, GPT, or Gemini.