Thomson Reuters spends $40M to train Thomson legal and tax LLM from open-weight models
- Thomson Reuters launched Thomson, its first in-house proprietary LLM, saying it spent $40 million on talent and compute after starting from an open-source foundation and retains full ownership and control of the model.
- The company trained Thomson with proprietary material from Westlaw, Practical Law, Checkpoint, and Reuters, while involving hundreds of subject-matter experts in training objectives and final evaluations.
- Thomson Reuters says Thomson has trained on less than 10% of its content so far and reports early evaluations put it on par with frontier models across a range of tasks, without publishing those results in the announcement.
- The company has begun using Thomson in CoCounsel Legal, plans sovereign AI options, and released a small open-weight version on Hugging Face for academic and non-commercial use.
- Thomson Reuters says external legal and AI academics are evaluating the model; two quoted law professors found its answers and citation quality competitive with ChatGPT, Claude, or leading frontier models in their tests.
Hacker News opinions
I could not find the technical report from the announcement. They do not seem eager to publish their evaluation results.
The Hugging Face release may be enough to run the evaluations independently. It includes Thomson-1.0-Small.
I do not see how Thomson Reuters earns back the $40M. This feels like a news agency chasing a technology trend to stay relevant.
Reuters news is only about 10% of Thomson Reuters revenue. Legal and tax products are much larger parts of the company.
I see this as a value add for existing data products and a moat against large LLM vendors. Bloomberg is a reasonable comparison because news is not its primary product either.
I doubt direct payback is the goal. Thomson Reuters can reduce exposure to model price increases, behavior changes, and vendor dependence, while possibly lowering inference costs for sustained internal use.
I expect more large organizations to train specialized models on data that frontier labs cannot access. The question is whether they can find a useful product and sell access to it.
Spending $40M for a marginal improvement surprises me. Why not use open-weight models directly?
The small model's Hugging Face page says it is based on Qwen3.6-35B-A3B. Business Insider also reported that Thomson was built on Qwen.
Calling citation quality "generally competitive" with frontier models is a weak pitch. Why pay for an extra product if frontier models already do the job?
I think it is mainly a cost-saving and marketing story.
Internal inference has a utilization problem. A model running at 30% GPU utilization is usually more expensive than an API, and I do not know how Reuters solves that long term.
The technical report describes continual learning rather than a simple small fine-tune. It says the small-model pipeline used about 1.63 × 10²³ FLOP and 35,207 B200 GPU-hours.
So it is still a Qwen fine-tune, which is reasonable. The press release makes the work sound more detached from the rest of the industry than it is.