Google Forges Diffusiongemma-26B-A4B-It A Diffusion Model That Denoises Entire Text Blocks At Once
Google has released a new open-weights AI model called Diffusiongemma-26B-A4B-it, which uses a unique method to generate text significantly faster than traditional models. Instead of creating text one token at a time, it uses a diffusion process to “denoise” large blocks of tokens all at once. This approach allows the multimodal model to process text, images, and video to produce text outputs.
Built by Google DeepMind that also brought us Gemma-4-12B-It and Gemma-4-12B, this model is part of the Gemma 4 family and uses a mixture-of-experts architecture for efficient performance. The design specifically targets high-speed generation, aiming to reduce the bottlenecks found in standard language models. It’s released as an open-weights model for the developer and research community.
Performance and efficiency highlights
Key features
- Generates over 1100 tokens per second on capable hardware.
- Uses block-autoregressive multi-canvas diffusion sampling for speed.
- Handles 256K token context windows for long documents.
- Built-in reasoning mode with configurable thinking capabilities.
- Supports 35+ languages and native function calling.
- Processes variable image resolutions with a visual token budget.
- Optimized for low-latency performance on a single accelerator.
- Encoder-decoder design separates prompt processing from text generation.
This tool is ideal for pro-consumer GPU owners and small agencies needing fast, local AI for tasks like document parsing, code generation, and chatbot development. Privacy-focused professionals can run powerful reasoning and visual understanding workflows without sending data to the cloud. The model’s speed and efficiency also make it a strong fit for serious hobbyists experimenting with advanced, local AI deployments.
What developers should know
While the model delivers impressive speed, its benchmark scores on reasoning tasks like AIME and GPQA Diamond are lower than the standard Gemma 4 26B model, reflecting a trade-off for faster generation. It is designed to work best with the recommended entropy-bound sampler, using a maximum of 48 denoising steps and adaptive stopping for confident predictions.
For optimal multimodal performance, users should place image content before text in prompts and select the appropriate visual token budget depending on whether they need speed or fine-grained detail.
"DiffusionGemma generates tokens using discrete diffusion. This open-weights model is multimodal, handling text, image, and video inputs to generate text output." — Source: Hugging Face
Project links
