Brian Armstrong: Coinbase fine-tunes Qwen, with fraud prevention outperforming cutting-edge models

Coinbase co-founder and CEO Brian Armstrong stated: The company has fine-tuned the open-source weight model Qwen3.5-9B using internal proprietary data, which performs better than cutting-edge models in the specific task of fraud prevention, while also being more cost-effective and faster.

Armstrong mentioned: Cutting-edge models are excellent general models, but companies will increasingly use specialized models for narrower tasks, which can operate independently or collaborate with teams. He believes this pattern will become more prevalent within the company.

He also noted that Coinbase is making it easier for employees to fine-tune models internally and providing relevant skills in internal tools, with example instructions for fine-tuning Qwen for payment risk audits. He anticipates that some backend services may be completely replaced by this new architecture.