A developer has released RustingStrata, a pure Rust inference engine for Qwen GGUF models (including MoE variants up to 80B) that compiles CUDA kernels at startup via NVRTC and supports Vulkan. The engine optimizes MoE models exceeding VRAM by offloading inactive experts to the CPU while concurrently executing GPU work, dynamically migrating frequently used experts back to the GPU. On an RTX 3060 (12 GB), the engine achieves approximately 1.7x faster decode performance compared to llama.cpp for Qwen models.