FS-DFM: New Apple AI models spit out long texts very quickly

Apple’s AI research department has presented a new approach to accelerating text generators. By means of so-called Few-Step Discrete Flow-Matching Language Models, FS-DFM for short, the throughput rate (“faster sampling and corresponding latency/throughput gains”) can be increased by up to 128 times, measured by the issuance of 1024 tokens. FS-DFM combines 1024 individual steps of conventional models in just eight steps, the scientists explain.

Technology from the field of image generation

In the Paper, which was published on the preprint server arXiv, it goes on to say that FS-DFM is supposed to create long texts particularly accurately. The diffusion technique is used, a generation process that was originally used primarily for image generation, but is now also used for large language models (LLMs) because it promises efficiency gains. Competing systems such as OpenAI’s GPT-5, on the other hand, rely on the Transformer technique, which uses so-called autoregression.

Diffusion models have the fundamental advantage that they can generate several output tokens, i.e. the individual components of the text output, at the same time. They are then further refined until it comes to the output to the user. Flow matching models are intended to further accelerate what Apple has now succeeded in doing – although existing processes are said to be outdated.

GPT-based LLMs are still mainstream

One problem remains that the selection of publicly available discrete flow matching models is still small – the AI scene continues to rely primarily on the GPT-based approach for text generators. Apple’s researchers want to help here by making their own DFM and FS-DFM models publicly available. Details can be found in the paper.

The project was developed in collaboration with researchers at Ohio State University. Apple’s AI research department had previously experimented with flow matching models in other areas, including for the Area of programming and in the Protein Folding. The company hopes that this will also lead to breakthroughs for its products. Apple still does not officially offer its own chatbot.