The Canadian data and legal giant Thomson Reuters wants to rely less on costly AI from Anthropic — and it turned to a Chinese open-source model to do it.
Thomson Reuters launched its first AI model, Thomson-1, on Monday, and it’s based on Snowdon, which was built by “realigning” an open-source Qwen model from the Chinese tech giant Alibaba, its chief technology officer Joel Hron said. Open-source means anyone can download the models for free and adjust them as they please.
Thomson-1 will take over some tasks previously handled by Claude, although Hron said that it’s not intended to replace the company’s work with Anthropic and other labs. Rather, Thomson-1 will focus on areas in which the company has expertise, starting with document review.
The shift to building an in-house model comes as large companies feel the pinch from high AI costs driven by chatbots like Claude and OpenAI’s Codex. It’s also a sign that cheap Chinese AI could be a boon to Western tech companies, with many startups already building products off Chinese open-source models.
Hron said Thomson Reuters decided to build its own model to better harness the company’s expertise in specific areas. He also cited cost as one of the main reasons.
Thomson Reuters expanded its partnership with Anthropic in May for its AI legal assistant, CoCounsel, which Hron said still relies mostly on Claude.
“Our main objective is to make Thomson the model that powers more and more of CoCounsel’s capabilities over time,” Hron told Business Insider.
The rise of Chinese AI models
Not everyone has embraced the rise of open-source Chinese models.
Anthropic has said that Chinese labs are illicitly “distilling” the outputs of its AI models to train their latest chatbots, and has called for the US to impose restrictions. Other government leaders, such as Sen. Tom Cotton, have raised concerns about US companies like Airbnb and Cursor using Chinese open-source models, citing potential security threats, including backdoors.
Anthropic and Alibaba didn’t respond to requests for comment.
Airbnb told Business Insider that it overwhelmingly uses US AI models, and that the few Chinese models it does use are open-source and run through US cloud providers. “Privacy and security come first in our approach to AI,” an Airbnb spokesperson said.
Hron said a joint team from Thomson Reuters and Imperial College in the UK adapted Qwen over several months to build Snowdon, ensuring it was “ethically and politically de-biased and safe to use.”
“There’s nothing that necessarily ties us to Qwen,” Hron said.
Having its own model means Thomson Reuters can build off its own intellectual property, rather than continually paying outside AI companies for theirs.
Hron compared it to renting versus 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.”
