Engineers Applu and Ohio State University introduced a new language model FS-DFM (Few-Step Discrete Flow-Matching), which according to a study can write long texts up to 128 times faster than conventional models of the type ChatGPT – while maintaining the same output quality. This makes it one of the most efficient text generation solutions ever released.
You might be interested in
How FS-DFM works
Unlike classic autoregressive models that generate text word by word (or “token”), FS-DFM uses the principle of diffusion models. These generate multiple words in parallel and gradually “improve” them in several steps. Apple but he went further and used variacalled flow-matching, which can speed up the entire process by skipping most iterations and generating the final text after just a few rounds of editing.
In the study Apple reports that FS-DFM can write a complete text in just eight fast iterations – while competing models require over a thousand steps. To achieve this, the research team took a three-phase approach:
- the model first learns to function with a different number of iterations,
- is subsequently "guided" by a larger teaching model, which helps him make more precise adjustments,
- and thenonec Apple optimized individual steps so that the model converges faster and more stably.
Speed without losing quality
In FS-DFM tests, it outperformed significantly larger diffusion models such as Dream (7 billion parameters) or LLaDA (8 billion parameters) in two key metrics – perplexity (naturalness of text) and entropy (stability of word selection). Despite its smaller size, it achieved lower perplexity values and more stable entropy, indicating higher accuracy and smoothness of output.
Gallery
Apple plans to make the code and model checkpoints available to allow researchers and developers to continue their research. The study “FS-DFM: Fast and Accurate Long Text Generation with Few-Step Diffusion Language Models” is already available on the arXiv server and offers detailed examples of how the model modifies words in the text in individual iterations.
If this approach proves successful, it could dramatically speed up text generation in applications such as Siri, Apple Intelligence or iWork, where both speed and accuracy of output become a key factor.
