Along with the model, Sber published tools that accelerate the training of diffusion models.

Sber introduced GFusion – an experimental diffusion language model developed based on the GigaChat neural network. A key feature of the project, created by an intern from the fundamental models team, was a fundamentally different approach to text generation, which opens up new opportunities for research in the field of artificial intelligence.

Unlike classic language models that generate text sequentially, word by word, GFusion works similarly to image generation. The model first forms a general “sketch” of the response and then refines it step by step. Sber noted that this allows text to be generated up to 45% faster compared to GigaChat 3 and provides greater flexibility: the model can supplement any part of the response without strictly following a left-to-right order.

In addition to the model itself, Sber published tools to accelerate the training of such systems and made changes to the popular SGLang framework, adding GFusion support to it. This is the first open-source project of this scale in Russia for creating diffusion language models.