Thomson Reuters Continues Training Qwen on 200B Tokens, Releases 35B Legal Model for Noncommercial Research

Thomson Reuters has introduced its first in-house Thomson model family and released the 35B-parameter Thomson-1.0-Small on Hugging Face. Rather than being pretrained from scratch, the model starts from Qwen3.6-35B-A3B and undergoes continued training from the intermediate Snowdon checkpoint. Its weights can be loaded with Transformers, vLLM, and SGLang, but they are licensed under PolyForm Strict 1.0.0, which permits only academic and noncommercial use; it therefore should not be regarded as a conventional open-source model.

According to the company, 200 billion tokens were selected for mid-training from a data pool containing more than 19 trillion tokens. The training mix consisted of three main components: proprietary documents, synthetic rewrites of those documents, and general-capability replay data intended to prevent catastrophic forgetting. Post-training additionally incorporated DPO, reinforcement learning, data derived from domain ontologies such as legal IRAC, and Deep Research trajectories designed to encourage faithful tool use and citations. The full process consumed approximately 35,207 B200 GPU-hours and 1.63×10²³ FLOP. The company says it invested $40 million in the overall model program.

Under the same evaluation pipeline, the Small model achieved an overall average score of 74.6, compared with 71.7 for the base Qwen model. Its Harvey Legal Agent Benchmark score rose from 69.5 to 73.4, while its PRBench Hard score increased from 26.9 to 31.4. The gains were not universal, however: its Terminal-Bench 2.1 score fell from 45.2 to 40.5, and its Humanity’s Last Exam score declined from 14.1 to 13.4. These results suggest that replay data can mitigate—but not eliminate—the trade-offs involved in domain-specific continued training.

What makes this release noteworthy is not the promotional claim that a “legal model surpasses frontier models,” but how an enterprise can convert private corpora into supervisory signals while retaining a compact, deployable MoE foundation. Engineering teams should next wait for independent reproductions, data-contamination analyses, and real-world tests of citation accuracy. For now, the reported scores, data recipe, and cost figures remain primarily vendor-supplied.