Thomson Reuters drops its own legal LLM inside CoCounsel

Training a domain-specific model doesn't require burning hundreds of millions on a foundation model from scratch.

Legal information giant Thomson Reuters just unveiled Thomson, its first proprietary large language model built specifically for legal work. Instead of trying to reinvent the wheel, the company took an open-weight base model and layered its proprietary legal content, professional expertise, and custom reinforcement learning on top.

The economics behind it are worth noting. While Thomson Reuters spent roughly $40 million over two years on computing and staffing for the project, the final training run cost just about $450,000.

Tapping 150 years of Westlaw

Thomson will debut as the default engine inside Tabular Analysis, a high-volume document review tool in the CoCounsel Legal AI assistant. However, CoCounsel isn't locking itself in; it remains a multi-model platform where administrators can switch back to third-party frontier models, and outside models will still handle tasks where Thomson doesn't hold an edge.

To build Thomson, the company relied on its crown jewel: the Westlaw platform, which holds over 40,000 individual databases and more than 150 years of curated legal editorial history. Interestingly, senior research scientist Andrew Bean noted that only about 10% of that total archive has been used so far.

Hundreds of legal experts helped guide the targeted post-training, crafting specialized prompts and grading responses in blind evaluations. The team also used reinforcement learning so the model learned how to query existing enterprise tools like Westlaw and Practical Law directly.

According to internal benchmarks, Thomson matches leading frontier models when relying only on web access, and reaches parity or slightly outperforms them on answer completeness and citation accuracy once connected to Thomson Reuters content. Academic and third-party testing is currently underway, with a smaller open-weight model heading to Hugging Face under a noncommercial academic license.

Why it matters

We think Thomson Reuters is executing the definitive blueprint for domain-specific enterprise AI.

Rather than competing with OpenAI or Anthropic on raw multi-modal reasoning across every discipline, TR focused entirely on where general labs cannot compete: curated proprietary data, verified citations, and reliable integration with industry-standard legal software. By keeping the final training run around $450,000, TR can easily swap in whatever open-weight base model leads the pack six months from now without re-spending tens of millions.

For corporate legal departments and law firms, document review is notoriously vulnerable to LLM hallucinations and fabricated citations. A model natively grounded in 150 years of Westlaw authority could dramatically reduce verification overhead, making it worth testing for repetitive contractual diligence.

The lingering question is pace. General-purpose frontier labs are expanding reasoning capabilities at a breakneck speed. If the next generation of frontier models crushes legal benchmarks simply through raw intellectual horsepower and tool-use reasoning, TR will have to continually justify the overhead of maintaining an internal model architecture. Furthermore, the performance claims still lack widespread independent validation, so outside verification of the upcoming Hugging Face weights will be the real test.