Coinbase CEO Brian Armstrong said the exchange fine-tuned the open-source model Qwen3.5-9B with internal data, achieving better fraud detection performance than frontier models at lower cost and faster speed. The model’s F1 score is 9.60% higher and end-to-end latency is reduced by 55.00%. Armstrong believes specialized models will win in niche tasks, and Coinbase is enabling employees to fine-tune models internally, with some backend services potentially replaced by this new architecture.