Thomson Reuters has launched its first proprietary large language model after spending about $40 million over two years on computing and specialist staff. The model, called Thomson, will initially power document-review work inside the company’s CoCounsel Legal product.
A large language model, or LLM, is the technology that interprets and generates text in systems such as AI assistants. Thomson Reuters is focusing its model on legal and other specialist work rather than trying to compete with the largest general-purpose systems across every subject.
The company said the $40 million investment covered talent and computing. It also said the final training run cost about $450,000, according to SiliconANGLE’s reporting from a company briefing.
Thomson Reuters owns Westlaw, Practical Law, Checkpoint and the Reuters news operation. Those collections give it legal, tax, accounting and news material that general AI developers cannot freely use in the same way.
Thomson was not built entirely from scratch
The word proprietary requires some explanation. Thomson Reuters owns and controls the specialized model it has produced, but the project began with an existing open-weight foundation rather than a blank system.
Open-weight models allow developers to obtain and modify the numerical parameters learned during training. That can avoid the expense of creating a foundation model from the beginning.
Thomson Reuters described its starting point only as a strong open-source foundation in its launch announcement. Chief Technology Officer Joel Hron subsequently identified it as a system derived from Alibaba’s Qwen 3.5, according to The Next Web.
Starting with Qwen does not, by itself, mean Thomson Reuters sends customer information or proprietary training material to Alibaba. The company says it controls Thomson and does not use customer data to train the model.
Thomson Reuters then used its own content, training methods and hundreds of subject-matter experts to adapt the foundation. The model has so far been trained on less than 10% of the company’s available content, according to management.
CoCounsel remains a multi-model product
Thomson’s first production job is Tabular Analysis, a CoCounsel Legal feature that reviews large collections of documents and organizes the results into a structured table. Thomson will be the default model for that task, although administrators can select alternatives.
The company is not abandoning outside AI suppliers. CoCounsel will remain a multi-model product, assigning different work to Thomson or third-party systems depending on which performs better for the task.
This limits the risk of tying the product to one model. It also gives Thomson Reuters a practical way to test whether its own system reduces operating costs or produces better legal work before extending it across the wider product range.
The company plans to add Thomson to more legal and tax products over the next year. It is also releasing a smaller open-weight version for academic and non-commercial use and opening the system to more outside evaluation.
Early performance claims still need wider testing
Thomson Reuters says its early evaluations put Thomson broadly alongside leading frontier models across several legal and general tasks. In company testing published in July, the model performed particularly well on instruction-following and long-context tests, while other systems led some reasoning and coding measures.
Those comparisons should not be treated as independent proof that Thomson is better than competing models. Some tests were designed or run internally, and the company said a fuller technical report and additional outside validation would follow.
The commercial case does not require Thomson to win every benchmark. Owning a specialized model can give Thomson Reuters more control over costs, product schedules and how its content is incorporated into AI services. CoCounsel can still use an outside model when that produces the stronger result.
Thomson Reuters said in February that CoCounsel had reached one million professionals across 107 countries and territories. The figure is company-reported and does not identify monthly active or separately paying users, but it shows the size of the product base into which Thomson can be introduced.
The launch follows a strong second quarter for the wider business. Thomson Reuters reported a 9% increase in total revenue and 8% organic growth, which excludes currency movements and acquisitions or disposals. It also raised its full-year revenue outlook.
The next test is operational: whether Thomson can lower the cost of serving CoCounsel customers while maintaining the accuracy, citations and privacy controls expected in legal work. That will matter more to the business than winning a single model ranking.
