Thomson Reuters has officially launched its proprietary large language model (LLM) named Thomson. Developed using the company’s own extensive database from Westlaw and Practical Law, Thomson is positioned as a specialized model intended to reshape the economics of professional artificial intelligence.
Joel Hron, Chief Technology Officer at Thomson Reuters, highlighted the innovative approach of crafting a domain-specific model with internal content, contrasting it with the prevalent trend in AI which emphasizes larger and more costly models like GPT. This effort, according to Hron, represents a shift towards building efficient models tailored for specialized areas such as legal research, which can operate at a lower cost yet maintain high performance.
The development of Thomson comes at an investment of $40 million, a fraction of the costs associated with the larger AI models on the market. The model’s training reportedly required only $450,000, showcasing the potential cost-efficiency of using proprietary data and expertise for training specialized models. Hron underscores that rather than starting from scratch, Thomson Reuters enhanced existing open-source models, like Qwen 3.5, to meet legal and tax-specific needs.
Thomson is set to be integrated first in CoCounsel Legal, enhancing features like Tabular Analysis for document review. This deployment strategy indicates Thomson’s utility in managing large volumes of structured data within legal contexts, while also maintaining the flexibility for administrators to choose other models if necessary.
The advent of Thomson does not only promise to optimize legal tools but also opens avenues for potential collaborations with law firms and corporations, who may fine-tune the model with their proprietary data for increased autonomy. Discussions are ongoing regarding licensing Thomson directly to these entities, as acknowledged by TR’s engagement with firms interested in expanding their in-house AI capabilities.
The process of building Thomson included utilizing the expertise of numerous subject-matter experts within the company, which involved highly specialized training on TR’s extensive proprietary content. The model achieves a balance between specific legal capabilities and general AI functions such as reasoning, math, and written language proficiency, as noted by Jonathan Schwarz, head of foundational research at Thomson Reuters.
Thomson Reuters is not merely focused on increasing the data input into Thomson but instead aims to expand the scope of its specialization through continued skill enhancements in legal and tax fields, leveraging its vast repository of proprietary content.
This strategic move by Thomson Reuters underscores its long-term commitment to integrating cutting-edge AI solutions within the professional sphere, promising substantial advancements in how legal professionals access and process information.