NEW YORK, August 24 — Thomson Reuters Corp. unveiled its first proprietary large language model, Thomson, on August 24, 2026, marking a strategic move to embed legal expertise directly into artificial intelligence workflows. The announcement, reported by Paul Gillin and published at 09:00 EDT, introduced a model built by adapting open-weight foundations with the company’s vast legal content, training methods, and professional knowledge.
Rather than constructing a model from scratch, Thomson Reuters combined external large language model technology with its own curated legal datasets, a process that reduced final training costs to approximately $450,000 after two years of investment totaling about $40 million in personnel and computing resources. Thomson will initially power Tabular Analysis, a high-volume document review function within the CoCounsel Legal AI assistant, serving as the default model for that specific task.
Administrators retain the ability to override the default and select alternative models when needed. Joel Hron, global head of artificial intelligence and TR Labs at Thomson Reuters, emphasized that the company’s goal is not to rival every frontier AI lab but to establish leadership in legal intelligence. “Thomson needs to set the frontier of intelligence for legal,” Hron stated, underscoring a distinct mission from broader AI research labs.
Andrew Bean, senior research scientist at the company, noted internal testing revealed Thomson performs competitively with leading models when relying solely on web-sourced information, and achieves superior accuracy when integrated with Thomson Reuters’ proprietary content. The model’s training involved realigning the base architecture with the company’s operational values, pretraining on internal legal publications, and reinforcement learning guided by subject-matter experts who evaluated blind comparisons of legal question responses.
Jonathan Schwartz, head of foundational research, highlighted that specialization risks eroding general capabilities, so the team prioritized continual learning to preserve breadth while adding domain depth. Schwartz explained that simply deploying an open-source model without Thomson Reuters’ refinements would lack alignment with company values and professional expertise. The training process incorporated curated legal datasets spanning over 40,000 Westlaw databases and 150 years of publishing history, with hundreds of experts contributing to objective evaluation frameworks.
Hron affirmed that Thomson’s design ensures it functions as a specialized tool within a broader multimodel ecosystem, using domain-specific strengths where advantageous while leveraging other models for complementary tasks. The launch represents a targeted application of AI to legal workflows, aiming to enhance precision and efficiency in document analysis for legal professionals.
