Thomson Reuters built its own AI model for ~$450K, pairing it with frontier models in production
bigdata · x · 2026-10-09
Thomson Reuters CTO Joel Hron explains on The Data Exchange why the company built its own specialized frontier model, Thomson, on open-weight bases plus decades of proprietary content.
Key points:
- Economics: $450K to train vs tens of millions for frontier models; started with basic fine-tuning, grew into a full training pipeline beginning from Qwen
- Data: proprietary data from 3,000 domain experts, selective pre-training, RL and agentic RL
- Production: inside CoCounsel, model orchestration and routing let the company swap base models and cut third-party reliance
- Trust: realignment, constitutional principles, human verification in tax workflows, continual learning and data privacy
The thesis: enterprises can pair open-weight models with proprietary domain data to get specialized capability at controllable cost, working alongside commercial frontier models.
Related event: Thomson Reuters Trains Own Model for $450K(2 posts)→
More from Companies & People
- Mistral CEO says it's 'a few months away' from catching up to frontier labs — WellsLucasSanto · 2026-10-09
- Perplexity's design team built its brand identity refresh tools on its own agent platform, Computer — AravSrinivas · 2026-10-09
- Princeton's Intro to Robotics planning & control module: videos, notes, code all free — Majumdar_Ani · 2026-10-09
- Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le found Discovery Loop to automate ML and science — JeffDean · 2026-10-09
- Charity Majors: Making your team happy is not a manager's job — business outcomes come first — rseroter · 2026-10-09
- "We bought Deno": developer world buzzing over terse acquisition announcement — michellechen · 2026-10-09