On August 24, Thomson Reuters, the content and technology company behind Westlaw, Practical Law, and Reuters, introduced Thomson, its first in-house developed large language model. The biggest departure from frontier AI labs: rather than pouring billions of dollars into compute over years, the company started from a strong open-source foundation and invested a total of $40 million across talent and compute.

How Thomson was built:

  • A strong open-source foundation, followed by modern mid-training and post-training techniques.
  • Training data drawn from decades of proprietary content: Westlaw, Practical Law, Checkpoint, Reuters.
  • Hundreds of domain experts involved from the design of training objectives through to final evaluation.
  • The company’s Fiduciary-Grade AI standard: customer data is never used for training without explicit consent.

According to the company’s internal evaluation, Thomson matches the latest frontier models across many general tasks. Compared specifically to the base model on which it was built, the model demonstrated marked improvements in two areas: following complex, multi-step instructions, and reasoning over dense, specialized documents. From this, the company drew a conclusion contrary to a widespread assumption: simply providing a powerful general-purpose model with access to the right content is not enough to achieve expert-level performance; dedicated training on proprietary data, combined with expert involvement during the training process itself, yields gains that simple retrieval cannot match.

The initial deployment is in Tabular Analysis within CoCounsel Legal, a high-volume structured document review workflow where a specialized model delivers an immediate edge. CoCounsel retains a multi-model architecture, deploying Thomson where it holds a clear advantage while keeping other models in place. The next plan is to expand across the company’s entire legal and tax portfolio, along with sovereign AI options.

Two details regarding transparency are worth noting. First, the company released a smaller open-weight version on Hugging Face for non-commercial academic research, while also having a group of legal and AI scholars evaluate it directly prior to launch. Second, a quote from Professor Jonathan Choi (Washington University) indicates he preferred Thomson’s answers over ChatGPT and Claude when testing challenging questions from his corporate tax class, thanks to citations linking to treatises that made the responses easier to verify. However, it should be noted: all current performance metrics are self-reported by the company, and a full technical report will be the place for independent verification.

The big picture: Thomson Reuters possesses something money cannot buy—a 175-year repository of specialized content and the trust of legal and tax practitioners. By building its own model, the calculation shifts from “which model is best” to “which model can be controlled and runs more cheaply.” If the company’s premise holds true, it represents a practical path for major data companies seeking to escape reliance on external AI labs, rather than entering an endless scaling race.