The 51.3-million-parameter model runs through the Cactus inference engine. Its low-bit format targets phones, robots and other memory-constrained hardware, with the Python package downloading the engine and model artifacts on first use.
Mirage is useful for applications that need to inspect images without routing every inference request to a remote service. It combines several visual tasks in one model so a developer can prototype extraction, detection and scene understanding through a shared interface.
