Brian Armstrong: Fine-tuned Qwen3.5-9B Outperforms Frontier Large Models in Fraud Prevention Tasks

Odaily reports: Coinbase CEO Brian Armstrong posted on X that after fine-tuning the open-weight model Qwen3.5-9B using Coinbase's internal proprietary data, the model surpassed frontier large models in the specific task of fraud prevention, while significantly reducing costs and running faster.

Armstrong believes that in the future, enterprises will increasingly adopt the approach of fine-tuning specialized models for specific tasks, and specialized models are expected to surpass general-purpose frontier models in certain niche tasks. Currently, Coinbase has integrated a model fine-tuning Skill into its internal AI tool framework, allowing employees to initiate the model fine-tuning and deployment process through simple instructions. He also predicts that in the future, some backend services may be completely replaced by this new architecture.