Thomson Reuters' 397B Qwen-based legal LLM beats GPT-5.4 on domain factuality
DeepLearningAI · x · 2026-09-09
Thomson Reuters launched Thomson, a proprietary LLM family built on Qwen with custom data engineering, targeting law, tax, finance, and news.
- MoE architecture: 397B parameters, 17B active per token, 262K-token context.
- Thomson-1.0-Large narrowly beats GPT-5.4 and Claude Sonnet 5 on completeness and factuality for tax/legal/news, and leads open-web factuality by 15 points; the Small model beats Gemma4-31B and Claude Haiku 4.5.
- Deploys first in CoCounsel Legal, business customers only.
- A 35B Thomson-1.0-Small will be released as open weights on Hugging Face for academic/non-commercial use.
Andrew Ng's DeepLearning.AI frames it as evidence that general-purpose AI isn't enough for regulated industries — specialized models retrained on proprietary data are an emerging enterprise paradigm.
More from Companies & People
- BlueDot's Frontier AI Governance Course: 10,000+ Alumni Placed in Policy Roles — AndyMasley · 2026-09-09
- Grok Bot product lead's playbook: run everything in the cloud, give each bot its own computer — threepointone · 2026-09-09
- Scoop: Altman Privately Opposes US Government Stake in OpenAI, Reversing Signals — ShakeelHashim · 2026-09-09
- Shipping AI Products Early Is Harder Than It Sounds — lucasmeijer · 2026-09-09
- Sam Altman says unprecedented Astra demand may force OpenAI to pause new Pro subscriptions — sama · 2026-09-09
- Hugging Face researcher to speak at Nerdearla in Argentina, side event planned — osanseviero · 2026-09-09