From rented intelligence to a model factory
For years Thomson Reuters offered AI-enhanced products by fine-tuning commercial models from third-party labs. Fine-tuning takes a pre-trained model and tweaks its weights on a smaller dataset, but it trims the model’s broad reasoning and ties the company to the provider’s pricing and API caps.
Now the company treats AI like a production line. In the past two years it hired engineers, data scientists, and compute specialists, and spent $40 million on cloud GPU time and on-prem hardware to train a model from the open-source Qwen series—Alibaba’s Qwen architecture. Open-source means the code and weights are public, so Thomson Reuters could start from a strong foundation without licensing fees.
A partnership with Imperial College produced an intermediate “Snowdon” model, first retrained for safety, ethics, and political neutrality—requirements any LLM must meet before reaching customers. After Snowdon, the team fed the model proprietary content from Westlaw, Practical Law, Checkpoint and Reuters news archives. Less than 10 % of that content has been used so far, leaving plenty of room to scale as more data are ingested. The final step added agentic reinforcement learning: the model interacts with Thomson Reuters’ own tool environments, receives feedback, and updates its policy to improve task performance. In this context, reinforcement learning teaches the model to achieve goals (like finding the right citation) through trial and error rather than static examples.
How the model stacks up
On Stanford LegalBench—a suite of legal-reasoning tests—the in-house model scored 0.823, behind Gemini 3.1 Pro, GPT-5.5 and Opus 4.8 on the Harvey Legal Agent Benchmark. It also lags on generic coding and reasoning problems, showing that its raw reasoning power remains weaker than frontier labs that pour billions into massive, general-purpose LLMs.
The advantage appears when the model draws on Thomson Reuters’ exclusive data. In a Deep Research benchmark measuring factual accuracy, the model hit 0.53 with only web access—lower than GPT-5.4’s 0.65. Adding its internal legal libraries pushed accuracy to 0.83, overtaking the commercial competitor. The result underscores a familiar pattern: a modestly sized model, when fed domain-specific knowledge, can beat far larger systems that lack that privileged information.
Why “own” beats “rent”
CTO Joel Hron and Research Chief Jonathan Schwartz point to three pillars for the investment:
- Cost and capability – Running high-volume tasks like document review on a commercial API incurs per-token fees that add up fast. A dedicated, smaller model processes the same workload at a fraction of the price while preserving legal-specific performance.
- Data sovereignty – Thomson Reuters’ most valuable asset is its curated legal content. Handing that data to an external AI provider creates security and confidentiality risks and hands the provider a competitive edge. Keeping the data in-house ensures improvements stay private.
- Compounding intellectual equity – Every time a lawyer reviews a model’s output, that interaction can be logged as training data, gradually refining the model. Over time the company builds a self-reinforcing asset that grows more valuable, like owning property rather than paying rent.
First use cases and an open-source nod
The model is rolling out in Thomson Reuters’ CoCounsel Legal suite for tasks that demand high throughput and low cost, such as the Tabular Analysis feature that extracts structured data from contracts. It also handles citation-checking, a repetitive but accuracy-critical step in legal drafting.
To nurture a developer ecosystem, a trimmed-down version will appear on Hugging Face under a non-commercial license. This lets researchers and partners experiment without exposing the full, proprietary model that powers the company’s revenue-generating products.