Quantized to Q4_K_M (4-bit) and served via llama.cpp. Expect ~5-10 tok/s. Quality is what you'd expect from a 0.5B model.