Tag
A user shares experiments running a 4-bit quantized DeepSeek v4 Flash on a 32GB M5 MacBook Air, achieving roughly 50 tokens/s prefill and 1 token/s decode using streamed experts and other tricks.