Thomson Reuters today (24 August) announced the launch of Thomson, the company’s first proprietary large language model, leveraging an open-source foundation and investing $40 million to train the model.
The launch represents the culmination of a strategy that has been quietly taking shape since Thomson Reuters acquired UK AI startup Safe Sign Technologies in 2024. At the time, the company said the acquisition would help accelerate the development of legal-specific AI capabilities within its CoCounsel platform, but few outside the business understood the scale of its ambitions.
Now Thomson Reuters is openly declaring that some of the most capable AI models no longer need to come from frontier AI labs.
Speaking as part of the launch, Thomson Reuters positions the new model as purpose-built for professional work, trained on decades of proprietary legal, tax, regulatory and news content and refined using extensive input from subject matter experts.
The company is not seeking to compete head-on with the likes of OpenAI, Anthropic or Google across every domain. Instead, its argument is that specialist intelligence will ultimately matter more than general intelligence when the task involves high-stakes professional work.
For legal professionals, the timing is significant. Over the past two years, law firms have increasingly adopted generative AI tools built on large general-purpose models. At the same time, questions have intensified around hallucinations, explainability, governance and whether generic AI systems can be trusted with complex legal workflows..
The model will initially power Tabular Analysis within CoCounsel Legal, a high-volume document review capability designed to process large numbers of legal documents in structured form. However, Thomson Reuters says it will continue to operate a multi-model strategy, deploying Thomson where it provides a clear advantage while continuing to utilise frontier models from providers including OpenAI and Anthropic elsewhere in the platform.
That approach acknowledges both the strengths and limitations of domain-specific models. Rather than arguing for a single-model future, Thomson Reuters appears to be betting on intelligent orchestration, matching the most appropriate model to the task at hand.
Here’s Joel Hron, Thomson Reuters’ chief technology officer, with more.
Why did Thomson Reuters build its own model?
“We built Thomson first and foremost for ourselves. We wanted more control over the intelligence inside our products: to make them better, improve the economics, reduce our dependence on someone else’s roadmap and give ourselves greater sovereignty over how that technology is trained and deployed. As access to powerful models becomes increasingly commoditised, we think owning more of that intelligence becomes strategically important.”
Why build on an open-weight model rather than continue relying entirely on frontier providers?
“Historically there has been a trade-off. You could use the most capable closed models and accept the dependency that comes with them, or you could take greater control with an open model and accept some distance from the frontier. What Thomson demonstrates is that those things do not have to be mutually exclusive. We can have control and sovereignty while still building a model that competes at a very high level.”
What makes Thomson difficult for someone else to replicate?
“The open-source foundation is the starting point, not the moat. The difficult part is everything that happens after that: the data, the training methodology, the domain expertise, the preference data and the evaluation infrastructure required to know whether the model is actually getting better at professional work. Thomson Reuters has spent decades building the corpus and expertise behind that system, and our research team has spent years working on how to translate those assets into model performance.”
Why did you only use about 10% of Thomson Reuters content?
“Less than 10% so far. And what comes next is not simply feeding the model more data. It is identifying which content will actually improve performance, turning that content into high-quality training data, and validating the results with the same expert-led process we have used to date. The fact that we have reached this level of performance using a relatively small portion of the content available to us gives us a lot of runway to keep improving Thomson across more tasks and workflows.”
What does Thomson mean for CoCounsel and Thomson Reuters products?
“Thomson gives us another form of intelligence that we control and can optimise around the work our customers actually do. We can use Thomson where it is the best model for the task, use third-party models where they are better, and keep evaluating that mix as the technology changes. The goal is not to force every problem through our model. It is to give CoCounsel the ability to use the right intelligence for the right work while improving the economics and control behind the system.”
Does Thomson mean customers have to choose between capability and sovereignty?
“That is one of the most important things we think Thomson demonstrates. A law firm or another professional organisation should not necessarily have to choose between the capability of a frontier model and the security, control and sovereignty advantages of owning more of its AI stack. We have shown there is a path to both. That creates a very different set of choices for organisations thinking seriously about their long-term AI strategy.”
Could law firms or other organisations eventually license Thomson directly?
“Yes. We built Thomson to make our own products better, but in doing that we have created a lot of optionality. There are law firms, corporate legal departments, development teams and other organisations that want to take more direct ownership of their AI strategy. They may want greater control over deployment, more sovereignty over their technology stack or the ability to build directly on a model designed for professional work. We are very open to working with those organisations and to Thomson being licensed directly.”
How big could that opportunity become?
“We built Thomson as infrastructure for Thomson Reuters. What we are beginning to see is that it could also become infrastructure for others. A lot of organisations today are deciding which models are best for which work. Thomson gives us the opportunity not only to be the company building applications on top of AI, but potentially to provide some of the underlying intelligence as well for companies making those model choices. That was not the primary reason we started the project, but it is a very interesting opportunity that comes from what the team has built.”
