Coinbase CEO Brian Armstrong announced that the exchange fine-tuned the open-weight Qwen3.5-9B model using proprietary internal data, achieving superior performance on specialized anti-fraud tasks compared to leading frontier models. The customized model delivered a 9.60% higher F1 score while reducing end-to-end latency by 55%, offering both cost and speed advantages. Armstrong noted that while frontier models excel as generalists, specialized models dominate granular applications. Coinbase has streamlined internal fine-tuning processes for employees and indicated that this new architecture may fully replace certain backend services in the future.