by Google
Google's open-weights 26B diffusion language model that drafts blocks of text in parallel.
Updated EU status checked Last shown
Sources: creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
DiffusionGemma is a diffusion-based language model from Google, described by creators as having 26B parameters. Instead of generating tokens strictly one after another as autoregressive models do, it drafts blocks of text in parallel and refines them, which is reported to give up to 4x faster generation.
Creators report that benchmark results are close to those of Gemma 4, and that it is released under the Apache 2.0 license. No official page, model card or paper was available when this entry was written, so these details come from a single creator summary and are unverified. Exact architecture, context length, and release date are not confirmed.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Open-source 26B diffusion language model that drafts text blocks in parallel, up to 4x faster than autoregressive models. Performs close to Gemma 4 on benchmarks; Apache 2 license.
Text diffusion model that drafts a 256-token paragraph at once, much faster than autoregressive models and roughly as capable as Gemma 4 12B. Aimed at efficient local use.
Closest neighbours by meaning, not just by category.
Model · Princeton
Fully open 3B text-to-image diffusion model from Princeton with released weights, training code, and data pipeline.
Model · DiffuSpace AI
Open-weights diffusion language model from DiffuSpace AI with 35B total and 3B active parameters.
Model · Amazon
Amazon's looped diffusion language models (1.7B and 8B) with adaptive per-token computation, built on Qwen3.
Model
Pixel-space text-to-image diffusion model transferred from latent diffusion models, without a VAE, supporting 4K output.
Model · Google
Google's open-weights 12B encoder-free multimodal model in the Gemma 4 family, taking text, image and audio input.
Model · Google
Google DeepMind's family of open-weights models built from Gemini technology, from phone-sized to laptop-class variants.