Coinbase CEO: Fine-tuned open-source small models outperform frontier models in fraud detection.

PANews October 12 news: Coinbase CEO Brian Armstrong posted on X that Coinbase used internal proprietary data to fine-tune the open-weight model Qwen3.5-9B, achieving better performance than frontier models on the specialized task of fraud prevention, at lower cost and faster speed. Coinbase's technical blog shows that the model's F1 score is 9.60% higher than that of frontier models, and end-to-end latency is reduced by 55.00%. Armstrong said that frontier models are excellent generalists, but specialized models will win on more narrowly defined tasks; Coinbase is making it easy for employees to fine-tune models internally, and some backend services may in the future be completely replaced by this new architecture.