Model · Multimodal language model

DiffusionGemma 26B A4B IT

Google DeepMind's ungated experimental open-weights multimodal model that generates text by discrete diffusion rather than token-by-token autoregression: an encoder-decoder over the 26B-total, 3.8B-active Gemma 4 mixture-of-experts architecture, denoising 256-token canvases with bidirectional attention for 256K-token context, released under Apache-2.0 and served on Vertex AI Model Garden.

Developer

Google

Distribution

Downloadable weights · Developer API

Artifact licensing

Open-licensed artifacts

License scope

Reviewed release artifacts and mandatory terms; not an assessment of training code or training data openness.

Licenses

Apache License 2.0

Boundary

This record represents the google/diffusiongemma-26B-A4B-it weight release published on Hugging Face at commit f7f5b7f5fa82ffc52addd066915886d497f5517b, the instruction-tuned 26B-total, 3.8B-active DiffusionGemma checkpoint. The Gemma 4 26B A4B autoregressive base it is fine-tuned from, the Diffusion Explained sample application, the community FP8 and 4-bit quantizations and the 31 finetune derivatives, and the Vertex AI Model Garden deployment of the same weights are separate boundaries.

Open in the directory → Open official model page ↗

Editorial ratings appear in the directory beside the profile they belong to and are never compared across collections.