Thomson Reuters just dropped their own LLM called Thomson, built on top of Alibaba's Qwen3.5-397B and Qwen3.6-35B. They didn't train from scratch — they took the open Qwen base, ran it through Imperial College London for safety/ethics/neutrality realignment (checkpoint: Snowdon), then hammered it with Westlaw, Practical Law, Checkpoint, and Reuters content. Domain experts labeled data and judged outputs. Final stage was agentic RL inside their own tooling.
$40M over two years. Final training run of the large version: ~$450K, three weeks. Team stayed under 36 engineers, peak hardware was 368 B200s. They've only tapped <10% of their content library so far. Customer data opt-in only.
CTO Joel Hron's logic: owning the weights means every expert correction becomes a permanent company asset. They're already swapping out Claude in production.
Benchmarks: Thomson-1-Large matches Claude Opus 4.8 on Stanford LegalBench, beats it on harder legal tasks and instruction-following, loses on coding and general reasoning. On their Deep Research test, it lagged GPT-class with web access alone but pulled ahead once it got Thomson Reuters' exclusive content. The other model improved nearly as much with the same data — so the edge is the proprietary archive and tooling, not just fine-tuning.
Smaller open-weight version hitting Hugging Face for academic/non-commercial use. Near Westlaw-grade performance. Blueprint is clear: grab a strong open model, specialize on proprietary text, then train it inside the actual products where users work.
$40M over two years. Final training run of the large version: ~$450K, three weeks. Team stayed under 36 engineers, peak hardware was 368 B200s. They've only tapped <10% of their content library so far. Customer data opt-in only.
CTO Joel Hron's logic: owning the weights means every expert correction becomes a permanent company asset. They're already swapping out Claude in production.
Benchmarks: Thomson-1-Large matches Claude Opus 4.8 on Stanford LegalBench, beats it on harder legal tasks and instruction-following, loses on coding and general reasoning. On their Deep Research test, it lagged GPT-class with web access alone but pulled ahead once it got Thomson Reuters' exclusive content. The other model improved nearly as much with the same data — so the edge is the proprietary archive and tooling, not just fine-tuning.
Smaller open-weight version hitting Hugging Face for academic/non-commercial use. Near Westlaw-grade performance. Blueprint is clear: grab a strong open model, specialize on proprietary text, then train it inside the actual products where users work.
