Thomson Reuters has built its own frontier models, aimed at legal, tax, and journalism. The open-weight Thomson-1.0-Small is based on Qwen3.6-35B-A3B and is on HuggingFace for academic and non-commercial use. Another version stays in-house.
- Thomson Reuters released Thomson-1.0-Small, a 35B parameter open-weight model built on Qwen3.6-35B-A3B and fine-tuned on decades of proprietary legal, tax, and news content, including case law, contracts, and regulatory filings.
- It outperforms Gemma 4-31B on several legal benchmarks and claims near-elimination of catastrophic forgetting through continual learning, though those claims rest on Thomson Reuters' own evaluation harness.
- A news and data company publishing an AI model aligned to something called the "Public AI Constitution" while using a PolyForm Strict license (not open source) is a tension worth noting.
