Running a 65B Coding Agent on a Single RTX 5090 Gaming Card

An author tested whether a 65B model could run a local coding agent on a single RTX 5090 gaming card. Using the Ripple-65B Preview model inside Hermes Agent with a small Python repository, the 3-bit IQ3_XXS file successfully passed all tests across multiple takes at about 41.8 tokens per second. Open article →