The GigaChat (Sber) team introduced GFusion — a diffusion language model (dLLM) developed based on GigaChat3-10B-A1.8B. Unlike standard autoregressive models, GFusion does not generate text one token at a time, but restores blocks in several passes using a diffusion mechanism.
What Happened
Developers presented the GFusion method, which provides an average generation speedup of 70% with a negligible loss in quality (within 2–4 percentage points). The technical implementation includes the use of custom kernels on TileLang to accelerate training and support in the SGLang inference engine using a new Entropy-Bounded Sampling algorithm.
Context
Traditional LLMs use an autoregressive approach, generating tokens strictly sequentially. GFusion was adapted from the existing strong model GigaChat3-10B-A1.8B, which allowed for the efficient implementation of a diffusion mechanism for parallel restoration of text blocks instead of classical character-by-character output.
Why It Matters for the Industry
Moving from sequential to parallel diffusion generation opens the way to a multi-fold increase in LLM throughput. The development of custom kernels and integration into the open-source SGLang ecosystem proves that optimization at the architecture and hardware levels can overcome the fundamental speed limitations of classical transformers and radically change the economics of AI products by reducing inference costs.
Why It Matters for Users
For end users, this means much more instantaneous interaction with neural networks. Instead of watching an answer being gradually "typed" letter by letter, models will be able to output entire paragraphs almost immediately while maintaining a high level of text coherence.
What Is Not Yet Known / Limitations
There is a difference in focus of evaluation: technical specialists (Architects) emphasize throughput and optimization, whereas business-oriented roles (Product Builders) focus more on changes in unit economics and user experience.
Sources
Author
Look at AI, Editorial Board