Thomson Reuters Launches Thomson, an In-House AI Model Built on Reworked Qwen Weights, to Cut Anthropic Reliance
Thomson Reuters spent $40 million building an in-house legal AI model on a reworked Alibaba Qwen base, aiming to reduce dependence on Anthropic and other outside AI labs.
Overview
Thomson Reuters launched Thomson, its first proprietary large language model, merging the company’s legal knowledge with external LLMs to deliver legal guidance, according to SiliconANGLE. Rather than building a foundation model from scratch, the company started with an open-weight model and layered on proprietary content, training methods, and professional expertise, an approach Thomson Reuters says is meant to reduce its reliance on outside AI providers including Anthropic, as reported by The Next Web.
What We Know
Thomson Reuters invested approximately $40 million across two years on personnel and computing for the project, though the final training run itself cost roughly $450,000, according to SiliconANGLE — a figure independently confirmed by The Decoder.
Thomson’s foundation is an intermediate model called “Snowdon,” built by reworking an open-source model from Alibaba, The Next Web and The Decoder both report. Thomson Reuters’ own launch announcement describes the starting point only as “a strong open-source foundation” without naming it; it was chief technology officer Joel Hron who identified the base as Alibaba’s Qwen in comments to Business Insider, as reported by The Next Web, which added that Business Insider identified the specific version as Qwen3.5. A joint team from Thomson Reuters and Imperial College London spent several months adapting Qwen, work Hron said ensured the result was “ethically and politically de-biased and safe to use,” per The Next Web.
On top of that base, Thomson Reuters applied pretraining on its own content, post-training guided by legal professionals, and reinforcement learning that integrated company tools including Westlaw and Practical Law, SiliconANGLE reports; hundreds of subject-matter experts helped define training objectives and evaluate the model’s responses. The Next Web reports the training content pulled from Westlaw, Practical Law, Checkpoint, and Reuters, and that only around 10% of the company’s overall content library has been used so far. Thomson Reuters said customer data is not used to train the model, per SiliconANGLE.
Thomson’s first deployment is inside Tabular Analysis, a document-review feature in the company’s CoCounsel Legal assistant, where it serves as the default model though administrators can choose alternatives, SiliconANGLE reports. CoCounsel remains “multi-model by design,” according to The Next Web, and Hron said CoCounsel still relies mostly on Claude — Thomson Reuters expanded its partnership with Anthropic in May for the same product, and Hron said the new model does not replace that work, per The Next Web. “Our main objective is to make Thomson the model that powers more and more of CoCounsel’s capabilities over time,” Hron told Business Insider, as reported by The Next Web.
Hron framed the strategy as an ownership question: “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,” according to The Next Web.
Chief executive Steve Hasker said “our early evaluations put Thomson on par with the latest frontier models across a range of tasks,” though that is the company’s own internal evaluation and the launch release did not publish figures alongside it, The Next Web notes. The Decoder reports company-published benchmark figures showing mixed results: on Stanford LegalBench, Thomson scores 0.823, trailing Gemini 3.1 Pro and GPT-5.5, while on the company’s own Deep Research benchmark, Thomson scores 0.53 for factual accuracy using only web access, versus 0.65 for GPT-5.4, but reaches 0.83 versus GPT-5.4’s 0.82 when given access to Thomson Reuters’ proprietary content.
Ahead of launch, Thomson Reuters shared the model with outside legal academics. Jonathan H. Choi of Washington University School of Law tested it against ChatGPT and Claude using questions from his Corporate Tax class and said: “All three models answered the questions correctly, but I preferred Thomson’s responses overall,” citing its links to treatises, per The Next Web. Professor Samuel Dahan, who directs the Queen’s Conflict Analytics Lab and the Cornell Legal AI Lab, found Thomson’s “citation quality generally competitive with leading frontier models,” including on Canadian employment-law questions, according to the same report.
Thomson Reuters is also releasing a smaller open-weight version of the model on Hugging Face under a noncommercial academic license, SiliconANGLE and The Decoder both report.
Thomson Reuters was not alone in the move: rival legal-AI vendor Harvey launched its own proprietary model, Tenet, a day earlier, post-trained on Kimi K3, the open-weight model from Chinese lab Moonshot AI, The Next Web reports. The Next Web also notes that Anthropic accused Chinese AI labs, including Alibaba, in June of illicitly distilling Claude’s outputs to train their own models, calling it the largest such distillation campaign it has identified.
What We Don’t Know
Thomson Reuters’ own launch materials do not name Qwen or confirm a specific version number; the Qwen and “Qwen3.5” identification traces only to Hron’s comments and Business Insider’s reporting as relayed by The Next Web, not to the company’s official release language. Independent, third-party verification of Thomson’s benchmark performance has not yet been published — the figures reported so far come from Thomson Reuters’ own internal evaluations, and the company said it is still in the process of sharing the model with outside academics and legal experts for external assessment.