Thomson Reuters built its own model on a Chinese open base

Thomson Reuters launched its first proprietary large language model on Monday. The company calls it Thomson. It says it spent $40mn on talent and compute to train the model, and that it began from what its announcement describes as “a strong open-source foundation”.

The announcement does not say which foundation. Its chief technology officer named it in an interview.

What the company says it spent

Thomson Reuters trades in Toronto and on the Nasdaq as TRI. It sells legal, tax, accounting and compliance products, and it owns Reuters. The announcement puts the training bill at $40mn, covering talent and compute.

It sets that against the frontier labs. Those labs have “typically spent billions of dollars on compute and years of infrastructure investment to reach the frontier”, the company says. It describes the result as a model it controls outright, “without the heavy inference costs of typical frontier models”.

The company has trained Thomson on less than 10% of its own content so far, according to the release. It names Westlaw, Practical Law, Checkpoint and Reuters as the sources.

What the base model is

Joel Hron, the chief technology officer, told Business Insider that Thomson starts from a model called Snowdon. Snowdon came out of “realigning” an open-source Qwen model from Alibaba. Business Insider identified the base as Qwen3.5, and reported the new model under the name Thomson-1.

The company’s own press release describes the starting point only as “a strong open-source foundation”. It does not name it.

Who built Snowdon

Hron said a joint team from Thomson Reuters and Imperial College in the UK adapted Qwen over several months. That work ensured the result was “ethically and politically de-biased and safe to use”, he said.

“There’s nothing that necessarily ties us to Qwen,” Hron told Business Insider. Alibaba said last month that it wants to charge its biggest users.

Where the model runs first

Thomson’s first deployment sits inside Tabular Analysis in CoCounsel Legal, the company’s AI assistant for lawyers. The release calls that “high-volume, structured document review”. It reaches law firms and corporate legal departments in the next release.

Thomson Reuters says CoCounsel Legal “remains multi-model by design”. It plans to extend Thomson across its legal and tax portfolio.

The iManage partnership four days earlier

Thomson Reuters and the document-management company iManage announced an expanded partnership on 20 August. It pushes CoCounsel Legal deeper into the iManage platform, alongside HighQ, Noetica and Legal Tracker.

The two companies also said they would add support for Model Context Protocol, so that approved Thomson Reuters tools can reason over iManage content while keeping access controls, ethical walls and privilege boundaries in place. Rawia Ashraf, co-head of CoCounsel Legal, said legal work “lives in too many places”.

What Claude still does

Hron said CoCounsel still relies mostly on Claude. Thomson Reuters expanded its partnership with Anthropic in May, for that same product.

“Our main objective is to make Thomson the model that powers more and more of CoCounsel’s capabilities over time,” Hron told Business Insider. He said the new model does not replace the company’s work with Anthropic and other labs.

How Hron described the decision

Hron cited cost as one of the main reasons. Owning a model lets the company build on its own intellectual property, he said, rather than paying outside AI companies for theirs. He compared it to renting against buying a house.

“Renting a house, you still have a roof over your head, and somebody’s taking care of it, and it’s great,” he said. “But you’re not building any equity that compounds into something valuable for you long term.” River AI raised $1.1bn this month to let companies train and keep their own models.

What Anthropic has said about Alibaba

Anthropic has accused Chinese labs of illicitly distilling the outputs of its models to train their own. In June it named Alibaba, in what it called the largest distillation campaign yet run against Claude. It has called for the US to impose restrictions.

Senator Tom Cotton has raised concerns about US companies using Chinese open-source models, among them Airbnb and Cursor. He cited potential security threats such as backdoors. Anthropic and Alibaba did not respond to Business Insider’s requests for comment.

The two academics quoted in the release

Thomson Reuters said it opened the model to legal and AI academics before launch. Its announcement quotes two of them.

Jonathan H. Choi of Washington University School of Law tested Thomson against ChatGPT and Claude, using questions from his Corporate Tax class. “All three models answered the questions correctly, but I preferred Thomson’s responses overall,” he said, citing the links to treatises.

Professor Samuel Dahan directs the Queen’s Conflict Analytics Lab and the Cornell Legal AI Lab. His evaluation found Thomson’s “citation quality generally competitive with leading frontier models”, including on Canadian employment-law questions.

The open-weight release

Thomson Reuters is publishing a “small” version of Thomson on Hugging Face as an open-weight model, for academic and non-commercial use. It says a technical report covering the foundation model’s development is available.

The same move at Harvey, a day earlier

Harvey launched Tenet on Sunday, its first proprietary model for legal work. Harvey post-trained it on Kimi K3, the open-weight model from the Chinese lab Moonshot AI. The desk covered Harvey’s in-house model that day.

Thomson Reuters cut engineering roles in July and said it was hiring AI-native staff instead. The desk covered those engineering layoffs at the time.

What the release claims about capability

Steve Hasker, the chief executive, said “our early evaluations put Thomson on par with the latest frontier models across a range of tasks”. That evaluation is the company’s own, and the release publishes no figures alongside it.

The release says Thomson shows “a meaningful uplift” over its base model in instruction following, and a greater uplift in navigating dense, domain-specific content. It says evaluations of the underlying foundation model appear in a technical report on the model’s development.

What Thomson Reuters promises customers

The release sets out a standard the company calls Fiduciary-Grade AI. It defines that as AI built for professionals with duties of care and accountability, “where almost right is not good enough”. The company says it does not use customer data to train the model without explicit consent.

Hundreds of subject matter experts worked on Thomson, according to the release, from the design of training objectives through to the final evaluations. Thomson Reuters says more “sovereign AI options” will follow.