Superpower Daily
Google and USC’s ME-POIs Adds Mobility Data to Place AI, Liftin...EverestLabs Puts AI Agents Near Recycling Controls, but Keeps O...Lexis+ with Protégé Adds Agents, Context and Office Files to Le...Seismic Completes Highspot Merger as Its Survey Finds 9% Embed...Geneva Rally Presses Hochul on 20-MW New-Project Permit Pause a...Trump Urges Cities to Welcome AI Data Centers as State Backlash...Google and USC’s ME-POIs Adds Mobility Data to Place AI, Liftin...EverestLabs Puts AI Agents Near Recycling Controls, but Keeps O...Lexis+ with Protégé Adds Agents, Context and Office Files to Le...Seismic Completes Highspot Merger as Its Survey Finds 9% Embed...Geneva Rally Presses Hochul on 20-MW New-Project Permit Pause a...Trump Urges Cities to Welcome AI Data Centers as State Backlash...
Modelspublished

Thomson Reuters Puts Its First Legal AI Model Into CoCounsel Document Review

Thomson Reuters will use its new legal model for specialized tasks while retaining third-party frontier models elsewhere in CoCounsel. Its performance claims have not yet received extensive independent validation.

By 3 min read
Thomson Reuters Puts Its First Legal AI Model Into CoCounsel Document Review
Sign in to save and react

Listen to this story

The audio brief

About 1:28
0:001:28
Read transcript
Thomson Reuters has put its first proprietary large language model, Thomson, into CoCounsel’s document-review workflow. Thomson is now the default engine for Tabular Analysis, a capability that helps analyze documents, but it is not taking over the entire product. CoCounsel remains multimodel: administrators can choose a different model for Tabular Analysis, while third-party frontier models continue handling other tasks. The strategy is specialization, not an attempt to build a general-purpose system from scratch. Thomson Reuters started with an open-weight base, then added its own legal content, training methods, and professional expertise. Hundreds of legal specialists helped define objectives, create example questions, and judge responses. Reinforcement learning also trained Thomson to work with Westlaw and Practical Law. The reported price was about forty million dollars over two years, including people and computing. Thomson Reuters says the final training run itself cost roughly four hundred and fifty thousand dollars. The company reports that Thomson was broadly competitive with leading models when all had web access, and roughly equal or slightly better when connected to Thomson Reuters content. But those results remain company-reported, without extensive independent validation. Customer data is excluded from training, and Thomson Reuters plans a smaller noncommercial version on Hugging Face, plus direct API testing. The key thing to watch is the promised technical report—and whether outside evaluations support the performance claims.

Story brief

3 key points

Thomson Reuters is positioning its legal model as a specialized layer inside a broader AI stack, rather than a replacement for frontier models. Thomson, adapted from an open-weight base and trained with Westlaw and Practical Law expertise, will power CoCounsel’s Tabular Analysis by default, while administrators retain model choice. The project cost about $40 million over two years, but its performance claims remain...

  1. 01

    The model was developed from an open-weight base, not trained from scratch.

  2. 02

    Thomson Reuters reported about $40 million in project costs and $450,000 for the final training run.

  3. 03

    CoCounsel remains multimodel; customers can override Thomson for Tabular Analysis.

Thomson Reuters has launched Thomson, its first proprietary large language model, and made it the default engine for Tabular Analysis, the document-review capability in its CoCounsel Legal AI assistant. The rollout puts a model trained on the company’s legal content and tools into a high-volume legal workflow.

Thomson is not replacing every model in CoCounsel. Thomson Reuters says the product will remain multimodel, using Thomson where legal specialization is advantageous and third-party frontier models for other work. Administrators can select another model for Tabular Analysis despite Thomson being the default.

Building on a base model, then narrowing the work

The company did not build a foundation model from scratch. It started with an open-weight model, then added proprietary content, training methods and professional expertise. Thomson Reuters says it is pursuing legal intelligence rather than competing with the largest AI labs across every field.

  • Pretraining added Thomson Reuters content to the base model.
  • Professionals guided post-training, including the training objectives and examples used for legal questions.
  • Reinforcement learning taught the model to work with Westlaw and Practical Law.

Thomson Reuters says hundreds of subject-matter experts helped define objectives, create legal-question examples and judge responses in blind comparisons. The tool-focused training is designed around the company’s existing legal research and practical-guidance products, not just a standalone chat response.

The reported build cost
About $40 millionProject spending

Thomson Reuters said people and computing for the project cost about $40 million over two years.

About $450,000Final training run

The company said the final training run cost about $450,000.

The performance case remains internal

Thomson Reuters says internal tests found Thomson broadly competitive with leading models when each had web access only. When connected to Thomson Reuters content, the company says it performed roughly equally or slightly better. The evaluations measured answer completeness and whether citations supported the claims made.

Those company-supplied results have not received extensive independent validation. Thomson Reuters expects to publish a technical report with additional benchmark results, leaving the external evidence for the performance comparison unfinished at launch.

Access plans broaden the test base

Thomson Reuters says customer data is not used to train Thomson. It is discussing direct model access with large law firms and corporations, and is open to customers adapting the model to their own knowledge and workflows.

The company has begun sharing Thomson with legal experts and academic institutions for testing. It also plans to release a smaller open-weight version on Hugging Face under a noncommercial academic license and is developing a portal for outside developers to request API keys and test the model directly.

Sources

  1. siliconangle.comThomson Reuters launches proprietary AI model for legal work - SiliconANGLE

Related coverage

Continue reading