35.107B total / approximately 3B active (MoE)
262,144 tokens natively
PolyForm Strict License 1.0.0 (noncommercial and restricted-use)
Official open-weight member of the Thomson-1.0 family, developed for specialized and general-purpose work with an emphasis on high-stakes professional domains including legal, tax, and journalism. Thomson-1.0-Small is a multimodal mixture-of-experts model derived from Qwen3.6-35B-A3B through the intermediate Snowdon1.1-Small checkpoint, with 35.107B total parameters and approximately 3B activated per token. Its Qwen3.5-MoE conditional-generation configuration retains image and video processing, uses 40 language layers, 256 routed experts with top-8 activation, hybrid linear and full attention, and a native 262,144-token context window. The development pipeline combines constitutional value re-alignment, continual pre-training, Direct Preference Optimisation, and reinforcement learning. Continual pre-training used 200B tokens curated from more than 19T tokens of permissively public and Thomson Reuters proprietary material, including news, contracts, regulatory filings, case law, statutes, and practitioner guidance, with replay data intended to preserve general capabilities. Post-training includes document- and ontology-driven preference data, practitioner queries, and agentic Deep Research training emphasizing faithful tool use and citation behavior. The model supports reasoning, structured tool calls, text generation, and image/video inputs through its supplied chat template and processor configuration. Thomson Reuters reports a 74.6 cross-domain benchmark average, including domain averages of 75.2 for legal and 82.6 for tax, alongside improvements over its Qwen3.6 base on document processing, human professional queries, deep research, factuality, and general-agent evaluations. Training consumed approximately 1.63e23 FLOP and 35,207 NVIDIA B200 GPU-hours. The weights are released under PolyForm Strict 1.0.0, which permits specified noncommercial uses but restricts commercial use, redistribution, modification, and derivative works.
https://huggingface.co/spaces/tri-fair-lab/publications/blob/main/Thomson_1_0_Technical_Report.pdf