Thomson Reuters Unveils $40M Proprietary AI Model Built on Alibaba's Qwen to Power Legal Work

Thomson Reuters launched Thomson, its first proprietary large language model, marking a strategic shift toward owning its AI infrastructure rather than relying solely on external providers like Anthropic1

1

. The company invested $40 million over two years in talent and compute to train the model, though the final training run cost just $450,000 thanks to efficiency gains3

3

. CEO Steve Hasker emphasized that Thomson proves what's possible when building AI on decades of proprietary content and editorial expertise, an advantage only Thomson Reuters has4

4

.

Built on Alibaba's Qwen Foundation with Ethical Safeguards

Thomson Reuters Chief Technology Officer Joel Hron revealed that the large language model starts from Snowdon, an intermediate model created by realigning Alibaba's Qwen open-source foundation1

1

. A joint team from Thomson Reuters and Imperial College London adapted the Qwen model over several months to ensure it was ethically and politically de-biased and safe to use2

2

. While the company's press release described only a strong open-source foundation without naming it, Hron confirmed to Business Insider that Qwen3.5 served as the base1

1

. Hron noted there's nothing that necessarily ties the company to Qwen, acknowledging that Alibaba announced plans last month to charge its biggest users1

1

.

Trained on Legal Content from Westlaw and Reuters

The training process incorporated Thomson Reuters' proprietary content, including data from Westlaw, Practical Law, Checkpoint, and Reuters1

1

. The company has trained Thomson on less than 10% of its total content so far1

1

. Westlaw alone encompasses over 40,000 individual databases and more than 150 years of legal publishing and editorial curation3

3

. Hundreds of subject-matter experts helped define training objectives, create examples of legal questions, and judge responses in blind comparisons3

3

. The training included realigning the base model with Thomson Reuters' values, pretraining on company content, targeted post-training guided by professionals, and reinforcement learning that taught the model to work with tools such as Westlaw and Practical Law3

3

.

Deployed First in CoCounsel Legal AI Assistant for Lawyers

Thomson's first deployment sits inside Tabular Analysis in CoCounsel Legal, the company's AI assistant for lawyers for high-volume, structured document review1

1

. The capability reaches law firms and corporate legal departments in the next release1

1

. CoCounsel Legal remains multi-model by design, with Hron confirming that the platform still relies mostly on Claude from Anthropic1

1

. Thomson Reuters expanded its partnership with Anthropic in May for CoCounsel1

1

. The company plans to extend Thomson across its legal and tax portfolio, with Hron stating the main objective is to make Thomson the model that powers more and more of CoCounsel's capabilities over time1

1

. By February, CoCounsel had been adopted by 1 million professionals across 107 countries4

4

.

Cost Efficiency and Intellectual Property Control Drive Strategy

Hron cited cost efficiency as one of the main reasons for building an in-house AI model, explaining that owning a model lets the company build on its own intellectual property rather than paying outside AI companies for theirs1

1

. He compared it to renting versus buying a house, noting that renting means you have a roof over your head and somebody's taking care of it, but you're not building any equity that compounds into something valuable long term1

1

. The company describes Thomson as a model it controls outright, without the heavy inference costs of typical frontier models1

1

. Thomson Reuters said customer data is not used to train the model without explicit consent, and controlling the model gives it more authority over deployment, governance, and future development3

3

.

Domain-Specific AI Competitive with Frontier Models

Internal tests showed Thomson is broadly competitive with leading models when all had access only to the web, moving to roughly equal or slightly better performance when connected to Thomson Reuters content3

3

. The tests assessed both the completeness of answers and whether citations supported their claims3

3

. Hron argued that Thomson doesn't need to compete with the largest AI labs across every field, stating that Thomson needs to set the frontier of intelligence for legal, which is a different job than what frontier labs are doing3

3

. The company plans to release a smaller open-weight version on Hugging Face under a noncommercial academic license and is developing a portal for outside developers to request API keys and test the model directly3

3

.

Expanded Integration with iManage and Sovereign AI Options

Thomson Reuters and document-management company iManage announced an expanded partnership on August 20, pushing CoCounsel Legal deeper into the iManage platform alongside HighQ, Noetica, and Legal Tracker1

1

. The companies also said they would add support for Model Context Protocol, enabling approved Thomson Reuters tools to reason over iManage content while maintaining access controls, ethical walls, and privilege boundaries1

1

. The company is discussing direct model access with large law firms and corporations and remains open to the possibility that customers could adapt Thomson to their own knowledge and workflows3

3

. Steve Hasker told the Financial Times in March 2024 that Thomson Reuters had amassed $8 billion to invest in the AI space and planned to expand in AI-driven professional services4

4

.