Google DeepMind ships DiffusionGemma, an experimental open model that writes text by diffusion

On June 10, 2026, Google DeepMind released DiffusionGemma, an open-weights language model that does not generate text one token at a time. It starts from a canvas of noise tokens in parallel, the same broad idea that powers image diffusion, applied to text. The pitch is speed: up to 4x faster generation on a GPU, with output that lands fast enough for interactive, local workflows.

AiPedia verified Google’s developer blog post and corroborating coverage on June 13, 2026.

What changed

DiffusionGemma is built on the Gemma 4 backbone, specifically the 26B-A4B architecture, and ships under a permissive Apache 2.0 license.

• Architecture: 26B Mixture of Experts with roughly 3.8B active parameters per inference step. It accepts interleaved text, image, and video input and returns text.
• Generation method: discrete text diffusion. Instead of left-to-right token prediction, it denoises blocks of 256 tokens at once, which is what unlocks the parallel speedup.
• Speed: Google cites 1000+ tokens per second on a single NVIDIA H100 and 700+ tokens per second on a consumer NVIDIA GeForce RTX 5090.
• Footprint: quantized, it fits within about 18GB of VRAM, which puts it inside high-end consumer GPU limits.
• Availability: free weights on Hugging Face, plus distribution through Google Cloud Vertex Model Garden and NVIDIA NIM.

Google is explicit that this is a research and developer release. It recommends standard Gemma 4 for production work that prioritizes output quality.

Why it matters

Almost every model buyers use today is autoregressive: it writes the next token, then the next, in sequence. That design caps how fast a single response can stream, because each token waits on the one before it. Diffusion-style text generation attacks that ceiling directly by producing many tokens in parallel.

For buyers and builders, the interesting signal is not that DiffusionGemma beats Gemma 4 on quality. It does not. The signal is that a frontier lab now ships a credible, open, local-runnable text-diffusion model that trades some quality for a large latency win, on hardware a developer can actually own.

That matters most for latency-sensitive local workflows: inline code completion, rapid edit-and-iterate loops, and generating structured or non-linear text where you want a draft instantly and will refine it. It is the kind of model that lives next to a coding agent, not the one that writes your final production answer.

Buyer action

Treat DiffusionGemma as a speed experiment, not a Gemma 4 replacement.

• If you run local models for Claude Code, Cursor, or Codex-style inline workflows, benchmark DiffusionGemma on your real latency-bound tasks and compare quality side by side against Gemma 4.
• Keep a quality-first model in the loop. Use the fast diffusion pass for drafts and interactive iteration, then escalate to a stronger model for anything that ships.
• Confirm the Apache 2.0 license fits your deployment before building on it. Permissive licensing is the practical reason this release is worth testing at all.
• Size hardware honestly. The headline tokens-per-second numbers assume an H100 or a 5090-class GPU; a weaker card will not reach them.

AiPedia verdict

DiffusionGemma is the most concrete sign yet that text diffusion is moving from research curiosity toward a real tool in the local stack. It will not replace your best model. It is a fast, open, GPU-friendly draft engine worth testing wherever latency, not final-answer quality, is the constraint. The fact that it is Apache 2.0 and runs on a single consumer GPU is what makes it more than a paper.