Thomson Reuters Builds In-House AI Model Thomson-1 Based on Alibaba's Open-Source Qwen to Reduce Dependence on Anthropic
Nashnova编辑部
Thomson Reuters launched Thomson-1, its first in-house AI model built on Alibaba's open-source Qwen, aiming to cut the steep cost of calling Anthropic's Claude — a move that signals big enterprises shifting from renting AI to owning it.
What exactly is Thomson-1?
Thomson Reuters' first in-house AI model, rebuilt on top of Alibaba's open-source large language model Qwen.
An intermediate model called Snowdon sits between Qwen and Thomson-1 — built over several months by a joint team from Thomson Reuters and Imperial College London, which "realigned" Qwen to ensure it is ethically and politically de-biased and safe to deploy.
In plain terms = they took a Chinese open-source model as a blank canvas, then re-fitted it with their own safety and compliance standards.
Why build a model in-house?
The core driver is cost. API fees for commercial models like Claude and OpenAI Codex have become a major line item in enterprise AI budgets.
CTO Joel Hron framed it as renting vs. buying a home. "Renting gives you a roof, and someone maintains it — that's great," he said. "But you're not building any long-term compounding asset."
This means → Thomson Reuters has concluded that perpetual reliance on third-party models is not just expensive — it accumulates no reusable technical capital. Every dollar spent leaves nothing behind.
Will Claude be fully replaced?
No — at least not soon. Thomson-1 will initially handle only a subset of tasks, focused on document review.
Thomson Reuters' AI legal assistant CoCounsel still runs primarily on Claude, and the company expanded its Anthropic partnership as recently as May this year.
Hron's exact words: "Our core goal is to have Thomson models progressively power more and more of CoCounsel's features."
This means → this is a gradual substitution strategy, not a hard switch — prove it in lower-risk workflows first, then push into the core product.
Does choosing Qwen create lock-in?
Hron stated explicitly: "There's nothing that necessarily locks us into Qwen," signaling the underlying model can be swapped in future.
This reflects a pragmatic stance: a key advantage of open-source models is replaceability — Qwen today, a better alternative tomorrow.
In plain terms = they are investing in the ability to customize, not in any single foundation model.
A Chinese open-source model — how does the West react?
Thomson Reuters is not alone — many Western startups and large companies now build on Chinese open-source models to avoid steep commercial licensing fees.
But the controversy is real: Anthropic has publicly accused Chinese AI labs of illegally "distilling" its model outputs for training and has called on the U.S. government to impose restrictions.
U.S. Senator Tom Cotton and other lawmakers have raised security concerns about companies like Airbnb and Cursor using Chinese open-source models, citing backdoor risks.
This means → cost savings and geopolitical risk are pulling in opposite directions — enterprises cut spending but may face political pressure and supply-chain security scrutiny.
Can this strategy actually work?
Two tests will decide: first, the cost-capability tradeoff — whether an in-house model truly saves money without sacrificing quality; second, precision — legal document review demands near-zero error tolerance, and Thomson-1 must match Claude's standard in that arena.
Neither Anthropic nor Alibaba responded to requests for comment.
In plain terms = the cost logic holds on paper, but the legal domain has no tolerance for mistakes — the real exam for Thomson-1 has not yet begun.
Content is for reference only, not financial advice.