Thomson Reuters: Thomson-1.0-Small (huggingface.co)

Thomson Reuters has unveiled Thomson-1.0-Small, an advanced open-weight frontier model designed for high-stakes applications in legal, tax, and journalism fields. Developed using a Continual Learning approach, this model builds on the open Qwen3.6-35B-A3B architecture and showcases significant enhancements across diverse performance domains. With its 35 billion parameters (3 billion activated), Thomson-1.0-Small addresses the economic impact of AI in professional settings, proving that robust frontier capabilities are attainable beyond major players in the industry. The model's development emphasizes important aspects of AI ethics and data usage. Through methodologies like Constitutional Direct Preference Optimization (DPO) and data-centric training on an extensive corpus of over 19 trillion tokens, Thomson-1.0-Small not only aligns with public values outlined in the Public AI Constitution but also curates proprietary knowledge while maintaining broad capabilities. Its results indicate a significant performance increase across multiple benchmarks, including legal reasoning and document processing. Moreover, the model's architecture is compatible with popular frameworks like Hugging Face Transformers, making it an accessible tool for the AI/ML community focused on enhancing productivity in high-stakes environments.