Brian Armstrong: Fine tuning open source model Qwen3.5-9B outperforms cutting-edge models in anti fraud performance

Coinbase CEO Brian Armstrong stated that Coinbase utilized internal data to fine tune the open-source model Qwen3.5-9B, which performed better than cutting-edge models in anti fraud tasks, with lower costs and faster speeds. The F1 score of this model is 9.60% higher than that of the frontier model, and the end-to-end latency is reduced by 55.00%. Brian Armstrong believes that dedicated models will win in segmented tasks, and Coinbase is allowing employees to easily fine tune models internally, with some backend services being replaced by this new architecture in the future.