Thomson-1.0-Small LLM Developed by Thomson Reuters (featherless.ai)

Thomson Reuters has announced the development of Thomson-1.0-Small, a cutting-edge 35.1 billion parameter Mixture-of-Experts (MoE) causal language model. Collaborating with Imperial College London, DatologyAI, and Lambda, this model is built on the Qwen3.6-35B-A3B architecture and supports a remarkable context length of 262,144 tokens. Utilizing a continual learning approach, Thomson-1.0-Small has been fine-tuned starting from the Snowdon1.1-Small checkpoint, enhancing its capabilities in both specialized and general-purpose domains, particularly for high-stakes professions like legal, tax, and journalism. This model is significant for the AI/ML community due to its emphasis on domain specialization and data-centric training, integrating proprietary information from over 19 trillion tokens, including news and regulatory filings. It addresses challenges such as catastrophic forgetting through continual learning, enhancing performance with an Overall Average score of 74.6% across multiple benchmarks. Notably, it achieves impressive results with 82.6% in tax-related tasks and 75.2% in the legal domain, while sustaining strong general performance. The model’s long context window and agentic capabilities make it well-suited for applications requiring accurate document analysis and question-answering in professional settings, elevating the standards for reliable AI tools.