There's lots of talk about companies "owning their own intelligence." Thomson Reuters is one of the first enterprises to pull it off.
On Monday, Thomson Reuters, the content and technology company, launched Thomson, its first proprietary large language model. Thomson was developed in-house, using an open-source model as its foundation and then trained on decades of proprietary content, including expertise from Westlaw, Practical Law, Checkpoint, and Reuters. The model is specifically designed for professional work and emphasizes the importance of domain expertise.
Because of the way the model was built, Thomson Reuters fully controls it, which gives the company a better understanding of the intelligence embedded in the model, reduces dependence on another company's AI roadmaps, and, most importantly, reduces costs. The company reports that the endeavor cost $40 million in talent and compute, which is comparatively low compared to what frontier labs spend on training.
"A real opportunity is for a company like Thomson Reuters to take those frontier open-source models and use them as starting points for our training and development with our unique data and compute. And in that way, we're able to train highly capable AI systems at a fraction of the cost that it would be to train one from scratch," said Joel Hron, Thomson Reuters CTO, in a briefing with the media.
So far, it has been trained on only 10% of Thomson Reuters' content, but the company says the goal isn't just to keep feeding it more content, but to turn more of it into targeted training data that improves the model.
The company reports that its early evaluations put Thomson on par with the latest frontier models across a range of tasks. It also attributes the quality of its content as the differentiator from proprietary models. When comparing how GPT-5.4, Sonnet 5, and Thomson 1.0 Large performed with web content only versus Thomson Reuters content only, there was, in every instance, some improvement in factuality or completeness. It also said that in its early evaluations, Thomson performed on par with the latest frontier models across a range of tasks.
At launch, it will be available in CoCounsel Legal, Thomson Reuters' legal AI platform built on its Fiduciary-Grade AI standard, which is designed to meet the requirements of highly regulated industries such as law, tax, and accounting.
Our Deeper View #
With the industry's focus on improving efficiency, there has been a strong emphasis on finding alternative ways to reduce costs when deploying models without compromising results. A key way to do so is to use domain-specific models, which are typically smaller and therefore require fewer tokens. They also tend to be faster and more accurate because they are working from a smaller, more specific data set. These domain-specific models are also key players in model routing, which promises to save users money by tapping into them when useful and is one of the industry's hottest trends. In Thomson Reuters' case, beyond saving money, a major advantage is being able to use its trove of data to inform the AI. However, the interesting thing is that the company has previously entered into licensing deals with AI labs such as Meta and Microsoft. We have to wonder if the company will take a different approach to content licensing now that it has its own proprietary model.