Have to say, Coinbase really embodies the typical approach of American companies in AI—constantly exploring different scenarios and experimenting with new models. Saw Brain’s tweet last week mentioning that Coinbase fine-tuned the open-source model Qwen3.5-9B using proprietary internal data for fraud prevention tasks. The precision, recall rate, F1 score, and dollar-weighted recall rate were all significantly better than the non-fine-tuned version and cutting-edge models like Opus 4.5. The fine-tuned 9B model also achieved a significant reduction in inference latency (P50 latency down by 55%), delivering higher speed and lower costs, showcasing the advantages of specialized models in enterprise-specific scenarios. Coinbase has internally developed simple tools (like the /fine-tune-and-serve-llm command) to help employees quickly fine-tune models. It’s expected that these tools could eventually replace some backend services and accelerate the adoption of AI applications within the company. Back in early July, it was Brain’s tweet https://(((x.com)))/brian_armstrong/status/2070670644577280109?s=46&t=k6rimWsEbo2D2tXolYcM-A that made me start paying serious attention to the progress of open-source models in enterprise AI adoption. This led me to propose the idea of enterprise AI adoption entering a “scenario-specific, multi-model” architecture, with the weight of open-source models rising rapidly (details in the quoted tweet). I also first discussed here https://(((x.com)))/qinbafrank/status/2074754779755295164?s=46&t=k6rimWsEbo2D2tXolYcM-A how the performance of large CSPs will become the most important metric for measuring AI commercialization. Just a couple of days ago, I revisited this topic here https://(((x.com)))/qinbafrank/status/2110568431192064077?s=46&t=k6rimWsEbo2D2tXolYcM-A, emphasizing again that the performance of large CSPs is the most comprehensive representation of AI commercialization.