Thomson Reuters launched its first proprietary LLM, called Thomson, spending $40mn on talent and compute. Its public announcement credits a vague 'strong open-source foundation,' but CTO Joel Hron told Business Insider the base is actually Alibaba's Qwen (Qwen3.5, internally realigned as 'Snowdon' with Imperial College). The model debuts in CoCounsel Legal's Tabular Analysis feature, though Claude still powers most of CoCounsel. The move mirrors Harvey's similar launch of Tenet, built on Moonshot AI's Kimi K3, a day earlier, amid ongoing scrutiny of US firms relying on Chinese open-weight models and Anthropic's accusations of Alibaba distilling Claude's outputs.
Table of contents
What the company says it spentWhat the base model isWho built SnowdonWhere the model runs firstThe iManage partnership four days earlierWhat Claude still doesHow Hron described the decisionWhat Anthropic has said about AlibabaThe two academics quoted in the releaseThe open-weight releaseThe same move at Harvey, a day earlierWhat the release claims about capabilityWhat Thomson Reuters promises customersQuestions this post answers
What open-source model did Thomson Reuters actually use to build its new Thomson AI model?
Thomson Reuters built its Thomson model on top of Alibaba's Qwen, specifically identified as Qwen3.5, though the company's own press release only described the base as 'a strong open-source foundation' without naming it. CTO Joel Hron confirmed the Qwen base and an internal derivative called Snowdon in an interview with Business Insider, after a joint team with Imperial College adapted it over several months for safety and bias mitigation. Track how legal AI vendors disclose their model foundations by following AI infrastructure news on daily.dev.
How much did Thomson Reuters spend to train its own Thomson AI model?
Thomson Reuters spent $40 million on talent and compute to train Thomson, its first proprietary large language model, trained so far on less than 10% of its own content from Westlaw, Practical Law, Checkpoint, and Reuters. The company contrasts this with frontier labs that have spent billions of dollars and years building infrastructure to reach comparable capability. Compare build-versus-buy AI cost tradeoffs like this one on daily.dev when weighing your own model strategy.
Is Harvey's new legal AI model Tenet also built on a Chinese open-source model like Thomson Reuters' Thomson?
Yes, Harvey launched its Tenet model a day before Thomson Reuters' announcement, post-training it on Kimi K3, the open-weight model from Chinese lab Moonshot AI. Both Thomson Reuters and Harvey built their first proprietary legal AI models on Chinese open-weight foundations within a two-day span, despite ongoing US political scrutiny over security risks in Chinese open-source models. Follow daily.dev for updates on which foundation models power competing legal AI products.