Coinbase says open-source AI fine-tuning improved fraud monitoring

Coinbase said it fine-tuned the open-weight AI model "Qwen3.5-9B" using internal data to block fraudulent transactions on its platform, producing results that outperformed frontier models. Coinbase said specialized models can deliver sharply better processing speed and cost efficiency than large general-purpose models. To support that effort, Coinbase said it built a simple command-based fine-tuning capability into its internal development environment, allowing employees to quickly develop AI models tailored to specific functions such as risk reviews. Based on the case, Coinbase said enterprise AI architectures are likely to shift from a reliance on general frontier models to combinations of task-specific models, and a substantial share of existing backend systems could eventually be replaced by that structure.