Apple and collaborators unveiled FS-DFM (Few-Step Discrete Flow-Matching), a new diffusion-style language model that dramatically accelerates long text generation. Unlike typical autoregressive models (which generate tokens one by one), FS-DFM generates multiple tokens in parallel and refines them through a small number of iterations.
The team showed that FS-DFM uses as few as eight iterative refinement steps to reach high quality outputs while maintaining strong performance in metrics such as perplexity and entropy.
The model’s speed advantage, up to 128× faster relative to competing diffusion models opens possibilities for more efficient and responsive generation of long-form content
