Thomson Reuters launches its own AI model after $40 million investment

Thomson Reuters is taking a major step forward in generative AI with the launch of Thomson, the company’s first proprietary large language model. The model was trained with a $40 million investment and will give Thomson Reuters greater control over its AI infrastructure, costs and data.

Thomson Reuters is now launching Thomson, the company's first in-house developed large language model, LLM. The investment means that the company moves from mainly integrating external AI models to also owning and controlling a central part of the AI technology to be used in the company's professional services.

Unlike many of the largest AI players, who have invested billions in computing capacity and infrastructure, Thomson Reuters has chosen a different strategy. Thomson builds on an established open source model that has then been further developed and specialized for the company's use cases.

• 

  OWN CONTENT

  
  

• 

  SPONSORED

  
  

• 

  ADVERTISEMENT

  
  

• 

  ADVERTISEMENT

  
  

In total, Thomson Reuters has invested 40 million dollars in training the model, including skills and computational capacity.

According to the company, the result is an AI model that Thomson Reuters fully owns and controls, while the costs of running the model are said to be significantly lower than for traditional frontier models.

Based on Thomson Reuters own data

An important part of the strategy is the company's extensive access to professional and industry-specific content.

Thomson has been developed with the support of decades of content, technology and domain expertise from Thomson Reuters. The work includes resources from Westlaw, Practical Law, Checkpoint and Reuters.

In addition, hundreds of subject matter experts have participated in the process, from the design of the training goals to the final evaluations of the model.

According to Thomson Reuters, less than ten percent of the company's content has been used to train the model so far.

The company therefore sees continued potential to develop Thomson through further specialization rather than simply increasing the amount of training data.

Challenging the pursuit of ever-larger AI models

The launch comes at a time when the AI industry has been focusing for several years on ever larger models, larger data sets and rapidly growing investments in computing capacity.

Thomson Reuters believes that Thomson shows that there is an alternative path.

By starting from a strong basic model and then specializing it for specific professional tasks, companies can, according to Thomson Reuters, create advanced AI at significantly lower costs.

Joel Hron, Chief Technology Officer at Thomson Reuters, describes the model as an example of how the economics of professional AI could change.

The company's early evaluations are also said to show that Thomson can perform on par with modern frontier models in several types of tasks.

AI sovereignty is becoming increasingly important

A central part of the launch also concerns AI sovereignty.

For companies and organizations, questions about how AI models are trained, where they run, what behaviors and biases are built in, and how information is protected are becoming increasingly important.

By owning the model itself, Thomson Reuters gains greater control over the entire chain.

It also reduces dependence on external AI vendors and gives the company greater control over how the technology is used in products where security, privacy, and accuracy are crucial.

Thomson Reuters particularly highlights the model's ability to follow complex instructions and work with extensive domain-specific material.

According to Thomson Reuters, the company's early results indicate that the combination of its own training data, human expertise and a strong basic model can produce results that cannot be achieved solely by giving a general AI model access to relevant information.

Independent experts test Thomson

Ahead of the launch, Thomson Reuters also made the model available to a group of researchers and experts in law and AI.

The purpose is to allow external actors to evaluate the model's performance and contribute to continued validation and development.

The company plans to continue making Thomson available for external evaluations in the coming months.

A smaller version of Thomson will also be made available as a open-weight model via Hugging Face for academic and non-commercial use.

First use occurs in CoCounsel Legal

The first commercial implementation of Thomson takes place in Tabular Analysis in CoCounsel Legal.

The function is used for structured analysis of large amounts of documents, an area where Thomson Reuters believes that a specialized AI model can provide clear benefits.

CoCounsel Legal will continue to use multiple AI models, with Thomson being used where its own model is deemed to provide the greatest benefit, while other leading models can be used for other types of tasks.

Thomson is expected to be available in Tabular Analysis for law firms and corporate legal departments in an upcoming release.

Thomson Reuters then plans to introduce the models within more parts of the company's legal and tax portfolio.

Own AI becomes a strategic asset

The launch illustrates a broader development in the enterprise market for AI.

As generative AI becomes an increasingly important part of mission-critical systems, competition is no longer just about which model is the biggest or most powerful. Control over data, training processes, costs, infrastructure, and model behavior is becoming increasingly important.

For Thomson Reuters, Thomson means that the company now controls another central part of its technology value chain.

The company already has the content, expertise and professional tools. With Thomson, the company also gets its own AI model that can be developed and specialized for future professional workflows.