Diffusion Language Models are Faster than Autoregressive on CPU: An Empirical Study of the Memory-Compute Regime Inversion

First empirical demonstration that diffusion language models outperform autoregressive models on commodity CPUs due to a memory-compute regime inversion. We present diffuse-cpp, the first C++ inference engine for diffusion LLMs with GGML quantization support, and an entropy-based adaptive scheduler that reduces denoising steps by up to 6x on easy prompts without quality loss.