Atomic Chat Adds DFlash Speculative Decoding, Claiming 2.2x Faster Local Qwen Inference

Atomic Chat has integrated DFlash, a block-diffusion speculative decoding mode, natively into its llama.cpp backend, promising up to 2.2x faster local inference on Qwen models with output the developers describe as byte-for-byte identical to standard generation. The feature is available across the app's macOS, Windows, and Linux builds.