Flyweight: open-source C++/CUDA engine for running MoE models bigger than your VRAM on one GPU + system RAM. First PyPI release, looking for contributors.

Flyweight is an open-source C++/CUDA engine designed to run Mixture-of-Experts (MoE) models, even those larger than a GPU's VRAM, by utilizing both GPU and system RAM. It supports various models like Qwen 3.x, DeepSeek-V4-Flash, and Gemma 4, and offers image input via mmproj on Qwen models. The engine also facilitates Z-Image-Turbo image generation alongside a chat model on the same card, with its first PyPI release now available.