Coinbase says it has fine-tuned a small open-weight artificial intelligence model that beats far larger frontier models at spotting payment fraud — and does the job more cheaply and far faster.
Chief executive Brian Armstrong announced on social media that the exchange post-trained the Qwen3.5-9B model on its own proprietary data for its Onramp fraud-detection work. According to the company’s engineering team, the adapted model scored 9.6 per cent higher on the F1 measure than frontier systems and cut end-to-end serving latency by 55 per cent. The announcement lands during a crowded week for technology news, as chipmakers Qualcomm and Arm await a jury verdict in their billion-dollar royalty battle.
Small, specialised and cheap
The work is described in detail in a three-part engineering series published by the company’s researchers. The team used reinforcement learning with verifiable rewards, scoring the model’s risk classifications against known fraud outcomes, and trained it on a single high-end graphics card for a total cost of under $100. In a live experiment, the AI-assisted payment flow recorded 30 per cent fewer fraudulent transactions and 22 per cent less fraud value than the existing system alone.
Armstrong argued the result points to a wider shift: frontier models are excellent generalists, but carefully specialised models will increasingly win on narrow, high-value tasks. Coinbase is now simplifying the fine-tuning process for its own staff, offering a one-click starting command in its internal tools.
Why it matters beyond crypto
Fraud is an unforgiving benchmark. A detector must catch rare criminal transactions without blocking legitimate customers, and it must do so in milliseconds at global scale. That Coinbase’s compact model outperformed systems many times its size suggests enterprises may get more from adapting open models to their own data than from renting the largest available system.
The company hints that some backend services could eventually be replaced by such specialised architectures altogether. If that proves true in production, the economics of applied AI may tilt further towards small, owned models — a notable turn for an industry that has spent two years racing to build ever-bigger ones.
