Thomson Reuters launches its own model based on Qwen to reduce reliance on Claude

Thomson Reuters (Thomson Reuters) launched Thomson-1, a language model developed internally on top of Alibaba’s Qwen3.5-397B, with the stated aim of reducing reliance on Anthropic’s Claude. The move signals a shift for the legal and financial information provider, which until now has depended on closed frontier labs for its core intelligence.

Thomson-1 is powered by Snowdon, a model co-developed by Thomson Reuters and Imperial College London. The team took the open-weight parameters of Qwen3.5-397B, the largest model in the Qwen3.5 series with 397 billion parameters, and performed a realignment to ensure it is “free of ethical and political biases”, according to the company.

The chief technology officer of Thomson Reuters explained that using frontier labs for intelligence is like renting a house: you get immediate access to a finished asset, but you have no control over the infrastructure and you depend on the owner for any change. Building on a model with open weights, by contrast, gives ownership of the base layer and the freedom to adapt it to specific needs without seeking permission.

It is important to note: Qwen3.5-397B is released with open weights, not full open source. The code for training, the data and the full methodology are not available, only the final weights. Thomson Reuters took those weights as a starting point and invested engineering and research resources in realignment, rather than starting from scratch or continuing to pay for access to a closed model.

The move illustrates a growing trend among organizations with sensitive data and regulatory requirements: they are willing to invest in internal engineering capability to control the base model, rather than remain dependent on an external supplier’s API. For Thomson Reuters, whose product is built on trust and legal accuracy, control over model biases and behavior is not only a technical issue, it is part of the business model.

The company has not published comparative benchmark performance against Claude or the base Qwen3.5-397B, and has not disclosed details of the de-biasing methodology beyond the general statement. Without independent data, it is difficult to assess whether the realignment improved the model’s fit for legal tasks, or harmed its general capabilities. The real test will be when Thomson-1 encounters real customers and cases.