What it is
DFlash is a small block diffusion model that drafts several tokens in parallel, which a larger target LLM then verifies, cutting generation latency. The project publishes draft checkpoints for model families such as Qwen, Gemma, Kimi, MiniMax, GPT-OSS and Llama, plus a Python package with a CLI for generation and benchmarking. It runs locally through Transformers or MLX on Apple Silicon, or against SGLang, vLLM and llama.cpp servers. People who serve open models and want faster decoding use it to pair a draft model with their target and measure the speedup on math and coding datasets.