Thomson Reuters introduced its first homegrown AI model, built by customizing Alibaba's open-source Qwen and training it on decades of the company's legal, tax, and news content. The two-year project cost $40 million in staff and computing, with the latest training run coming in at $450,000.
The numbers that matter
Thomson benchmarked the model ahead of Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5 in certain tasks, though testing has been internal only. CTO Joel Hron compared renting AI to renting a house, telling Business Insider you end up with no equity that compounds into something valuable over time. The company owns Westlaw, Practical Law, and Reuters, giving it proprietary content general AI labs cannot freely use.
An open-weights version of Thomson will be available on Hugging Face for researchers. Training so far has covered under ten percent of the company's overall content library. The model will power document review in CoCounsel Legal, with administrators able to select alternatives if needed.
Why this breaks the pattern
Forty million might sound like a lot, but to a giant company on expensive API plans, it's closer to a year of invoices than a moonshot budget. Expect more organizations with deep content libraries to run the same math, especially as training costs drop and strong open base models become the norm. Thomson Reuters proved you don't need billions to own your intelligence. You just need to stop believing only frontier labs can build it.