DiffusionGemma: 4x faster text generation

DiffusionGemma is an experimental open-source model that uses text diffusion to generate content up to four times faster than traditional autoregressive models. Unlike standard models that produce text one token at a time, DiffusionGemma generates entire blocks of 256 tokens simultaneously. This approach shifts the processing bottleneck from memory bandwidth to compute, allowing the model to utilize hardware more efficiently during local, single-user tasks.

The model is a 26B Mixture of Experts architecture that activates only 3.8B parameters during inference, enabling it to run on high-end consumer GPUs with 18GB of VRAM when quantized. Because it uses bi-directional attention, the model can evaluate and refine entire text blocks at once. This makes it particularly effective for non-linear tasks like code infilling, mathematical graphs, and in-line editing, where tokens must relate to one another regardless of their position in a sequence.

While DiffusionGemma offers significant speed advantages for local workflows, it is not intended to replace standard autoregressive models for all use cases. Its output quality is generally lower than that of the standard Gemma 4 family, and its performance benefits diminish in high-concurrency cloud environments. Released under an Apache 2.0 license, the model is designed for developers and researchers to explore rapid iteration and real-time interactive applications. It is compatible with major development tools and optimized for both consumer and enterprise NVIDIA hardware.