thomsonreuters/Thomson-1.0-Small
Thomson-1.0-Small is a 35.1 billion parameter Mixture-of-Experts causal language model developed by Thomson Reuters, built upon the Qwen3.6-35B-A3B architecture. It is designed for high-stakes professional work across legal, tax, and journalism domains, leveraging a continual learning paradigm and a 262,144 token context length. The model demonstrates strong performance in domain-specific tasks like legal document processing, tax Q&A, and deep research, while maintaining general capabilities.
Loading preview...
Model Overview
Thomson-1.0-Small is a 35.1 billion parameter Mixture-of-Experts (MoE) causal language model developed by Thomson Reuters, in partnership with Imperial College London, DatologyAI, and Lambda. It is built on the Qwen3.6-35B-A3B architecture and features a native context length of 262,144 tokens. The model was developed using a continual learning approach, starting from the Snowdon1.1-Small checkpoint, to achieve high proficiency in specialized and general-purpose domains.
Key Differentiators
- Domain Specialization: Enhanced for high-stakes professional work in legal, tax, and journalism, combining technical rigor with nuanced understanding.
- Continual Learning: Employs a pipeline that delivers significant improvements across capabilities while minimizing catastrophic forgetting, unlike limited fine-tuning or prompt engineering.
- Value Sovereignty: Aligned with the Public AI Constitution through Constitutional DPO and reinforcement learning, promoting transparent and scrutinizable normative foundations.
- Data-Centric Training: Incorporates proprietary Thomson Reuters data, including news, contracts, and regulatory filings, curated from over 19T tokens.
- Agentic Deep Research: Features a research harness with reward structures to reduce hallucinations and ensure faithful tool use and accurate citation in high-stakes environments.
Performance Highlights
Thomson-1.0-Small achieves an Overall Average score of 74.6% across various benchmarks. It shows strong performance in:
- Legal Domain: Achieves 75.2% average, with 73.4% on Harvey Legal Agent Benchmarks and 78.8% in Document Processing & RAG.
- Tax Domain: Scores 82.6% average, including 86.6% on Tax Q&A and 78.6% on Deep Research.
- General Capabilities: Maintains strong general performance, with 81.0% on Writing and 85.8% on General Agent tasks, while also demonstrating 98.5% political neutrality.
When to Use This Model
Thomson-1.0-Small is ideal for applications requiring high accuracy and specialized knowledge in legal, tax, and journalism fields. Its long context window and agentic capabilities make it suitable for complex document analysis, research, and question-answering in professional settings where factual accuracy and robust reasoning are paramount.