Google DeepMind has presented DiffusionGemma, an experimental artificial intelligence model that breaks with the traditional paradigm of sequential text generation. Instead of predicting word by word as conventional autoregressive models do (ChatGPT, Claude, Gemini), DiffusionGemma “draws” the entire text and then refines it progressively, achieving speeds up to 4 times higher in generation on GPU.
How it works
Traditional language models build responses linearly: they predict one word, then the next, and so on until the idea is complete. DiffusionGemma applies a diffusion logic, similar to that used by image generation models such as Midjourney or Stable Diffusion. The system starts from a global structure and progressively adjusts the entire content until it reaches a coherent version.
This means the model can generate complete blocks of 256 tokens in parallel at each step, using bidirectional attention that allows each token to “see” all the others simultaneously. The result is not only faster, but also offers significant advantages for non-linear domains such as in-line editing (infill) and code completion.
Technical specifications
DiffusionGemma is a Mixture of Experts (MoE) model with 26B total parameters, although it only activates 3.8B parameters during inference, which allows it to run on consumer GPUs such as the NVIDIA RTX 5090 or 4090 (quantized, it fits in 18 GB of VRAM).
- More than 1,000 tokens per second on an NVIDIA H100 GPU
- More than 700 tokens per second on an NVIDIA GeForce RTX 5090
- Up to 5× faster than comparable autoregressive models on Blackwell GPUs thanks to native NVFP4 support (4-bit floating point format)
Open source and accessible
Google has released DiffusionGemma under the Apache 2.0 license, which means any developer can download the weights, modify them and deploy them in their own projects. The weights are available on Hugging Face, and the model can run with frameworks such as Hugging Face Transformers, llama.cpp, MLX, Unsloth and NVIDIA NeMo.
In addition, it is a multimodal model: it accepts text, image and video inputs to generate text as output. It is designed for local interactive workflows where speed is critical: in-line editing, rapid iteration and non-linear text structures.
A paradigm shift
DiffusionGemma does not replace traditional autoregressive models. Google recommends Gemma 4 for production applications where quality is the priority, but DiffusionGemma opens the door to a new generation of diffusion-based language models that prioritize speed without sacrificing too much accuracy.
This approach, combined with increasingly powerful and affordable hardware, brings closer the dream of having truly interactive AI assistants running entirely locally, without depending on external servers or an internet connection.
Sources: Google Blog, Google DeepMind, Google AI for Developers.
