Thomson Reuters spends $40M building its own legal model on Alibaba's Qwen
Thomson Reuters fine-tuned Alibaba's Qwen3.5-397B into 'Thomson,' a legal model that scores 0.823 on Stanford LegalBench — behind Gemini 3.1 Pro and GPT-5.5 — and edges past GPT-5.4 on its own deep-research benchmark only once given Westlaw content, 0.83 to 0.82. The lead comes from proprietary data rather than the training: GPT-5.4 improves nearly as much with the same access, so the moat is exclusive content plus hundreds of in-house domain experts, and under 10% of that content has been used so far.
Source: the-decoder.com ↗
Every expert review during a product update becomes training data. With an in-house model, you are building equity in something that you own for the long-term, and that compounds over time.
Joel Hron, Thomson Reuters CTO
Why this matters
- → Proprietary data, not model training, drives the competitive edge — GPT-5.4 improves just as sharply with the
- → Shows specialized in-house models can compete with frontier labs for $40M, enabling independence from OpenAI/A
- → Demonstrates the moat is decades of exclusive content plus domain expertise, not the model itself
Owning the moat