BlockBeats news, on October 12, Coinbase CEO Brian Armstrong posted on social media that the company fine-tuned the open-source model Qwen3.5-9B using proprietary internal data, achieving superior performance on fraud detection tasks compared to state-of-the-art large models, while also being more cost-effective and faster. Relevant charts show that the post-trained model significantly outperformed both the unadapted version and the Opus 4.5 decision agent across metrics including precision, recall, F1 score, and amount-weighted recall. Online inference median latency was reduced by approximately 55% compared to Opus 4.5.
Armstrong believes that while frontier models excel as generalists, fine-tuned specialized models will increasingly outperform them on narrower tasks, whether operating independently or collaboratively. Coinbase is simplifying the fine-tuning process for internal employees and has already integrated one-click launch commands into internal tools; in the future, some backend services may be entirely replaced by such architectures.