# DiffusionGemma

> Google's first open-weight text-diffusion model: a 26B Gemma 4 MoE that denoises text in parallel for up to 4x faster generation.

- Maker: Google
- Line: Gemini Diffusion / DiffusionGemma
- Canonical page: https://ai-tldr.dev/models/diffusion-gemma/
- Updated: 2026-06-18

## In short

DiffusionGemma is an experimental open-weight large language model from Google, released on June 10, 2026 as part of the Gemini Diffusion / DiffusionGemma line. It has a 256K-token context window. You can use it through Hugging Face, NVIDIA NIM and Vertex AI Model Garden.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 256K (max output 256K) | [1] |
| Input price | Free (open weights) | [2] |
| Licence | Apache-2.0 · open weights | [1] |
| Released | 10 June 2026 | [3] |
| Parameters | 26B total (25.2B) MoE · 3.8B active | [1] |
| Knowledge cutoff | Jan 2025 | [1] |

## Sources

1. [DiffusionGemma model overview — Google AI for Developers](https://ai.google.dev/gemma/docs/diffusiongemma)
2. [google/diffusiongemma-26B-A4B-it — Hugging Face model card](https://huggingface.co/google/diffusiongemma-26B-A4B-it)
3. [DiffusionGemma: 4x faster text generation (Google blog)](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/)

## Overview

DiffusionGemma is an experimental open-weight large language model from Google, released on June 10, 2026 as part of the Gemini Diffusion / DiffusionGemma line. It is the first openly released text-diffusion model in the Gemma family: instead of predicting one token at a time like a normal autoregressive LLM, it starts from a canvas of placeholder tokens and iteratively denoises a 256-token block in parallel, the same idea behind AI image generators applied to text.

Under the hood DiffusionGemma reuses the Gemma 4 26B-A4B Mixture-of-Experts backbone (26B total parameters, about 3.8B active, 8 of 128 experts) with a new diffusion head and bidirectional attention. The instruction-tuned weights ship as google/diffusiongemma-26B-A4B-it under the Apache 2.0 license on Hugging Face, Kaggle, and Vertex AI Model Garden, support a 256K-token context window, and accept text, image, and video input (text output only; no audio).

The headline benefit is speed: DiffusionGemma reaches more than 1,000 tokens per second on a single NVIDIA H100 and 700+ tokens per second on an RTX 5090, up to roughly 4x faster than comparable autoregressive Gemma models, and fits in about 18GB of VRAM once quantized. The trade-off is quality: Google states its overall output quality is below standard Gemma 4, so it positions DiffusionGemma for speed-critical, interactive, local workflows rather than maximum-quality production use.

## Benchmarks

- MMLU-Pro: 77.6% ([source](https://huggingface.co/google/diffusiongemma-26B-A4B-it))
- GPQA Diamond: 73.2% ([source](https://huggingface.co/google/diffusiongemma-26B-A4B-it))
- AIME 2026 (no tools): 69.1% ([source](https://huggingface.co/google/diffusiongemma-26B-A4B-it))
- LiveCodeBench v6: 69.1% ([source](https://huggingface.co/google/diffusiongemma-26B-A4B-it))
- MATH-Vision: 70.5% ([source](https://huggingface.co/google/diffusiongemma-26B-A4B-it))
- MMMU Pro: 54.3% ([source](https://huggingface.co/google/diffusiongemma-26B-A4B-it))

## Pricing

- Input: Free (open weights) / 1M tokens

DiffusionGemma is released as open weights under Apache 2.0 — there is no per-token list price from Google; you self-host (Hugging Face / Kaggle / Vertex AI) or use a hosted endpoint such as NVIDIA's free NIM cloud API. Hosting and GPU costs depend on the provider.

Source: https://huggingface.co/google/diffusiongemma-26B-A4B-it

## Strengths

- First open-weight (Apache 2.0) text-diffusion model in the Gemma family, downloadable from Hugging Face, Kaggle, and Vertex AI
- Very high throughput — 1,000+ tokens/sec on an H100 and 700+ on an RTX 5090, up to ~4x faster than autoregressive Gemma
- Efficient 26B Mixture-of-Experts design with only ~3.8B active parameters; fits in ~18GB VRAM quantized for single-GPU local use
- 256K-token context window with text, image, and video input
- Day-one tooling support across vLLM (first diffusion LLM in the framework), Transformers, MLX, SGLang, and NVIDIA NIM

## Best for

- Reach for it when generation latency matters more than top-tier quality — interactive local chat, autocomplete, or streaming UIs.
- Reach for it to run a fast text model on a single consumer or data-center GPU after quantization.
- Reach for it to experiment with text-diffusion decoding (parallel block denoising) on open Apache 2.0 weights.
- Use standard Gemma 4 instead when you need maximum output quality for production-grade tasks.

## How to access

- [Hugging Face](https://huggingface.co/google/diffusiongemma-26B-A4B-it) — model id `google/diffusiongemma-26B-A4B-it`
- [NVIDIA NIM](https://build.nvidia.com/google/diffusiongemma-26b-a4b-it/modelcard) — model id `google/diffusiongemma-26b-a4b-it`
- [Vertex AI Model Garden](https://deepmind.google/models/gemma/diffusiongemma/)

## Gemini Diffusion / DiffusionGemma versions

- DiffusionGemma (10 June 2026) — this page
- [Gemini Diffusion](https://ai-tldr.dev/models/gemini-diffusion/) (20 May 2025)

## FAQ

### What is DiffusionGemma?

DiffusionGemma is Google's first open-weight text-diffusion language model, released June 10, 2026. It is built on the Gemma 4 26B-A4B Mixture-of-Experts backbone (26B total parameters, ~3.8B active) and generates text by denoising a 256-token block in parallel rather than one token at a time, which makes generation up to 4x faster than comparable autoregressive Gemma models.

### How fast is DiffusionGemma?

Google reports more than 1,000 tokens per second on a single NVIDIA H100 and 700+ tokens per second on an RTX 5090 — up to roughly 4x the throughput of comparable autoregressive Gemma models. The model fits in about 18GB of VRAM after quantization, so it can run on a single high-end GPU.

### Is DiffusionGemma open source and free?

Yes. The instruction-tuned weights (google/diffusiongemma-26B-A4B-it) are released under the Apache 2.0 license and can be downloaded from Hugging Face, Kaggle, and Vertex AI Model Garden. There is no per-token price from Google; you self-host or use a hosted endpoint such as NVIDIA's free NIM cloud API, paying only for the underlying compute.

### Is DiffusionGemma better than Gemma 4?

Not in quality. Google states DiffusionGemma's overall output quality is lower than standard Gemma 4 and recommends it for speed-critical, interactive, local workflows. For maximum quality, Google still recommends the autoregressive Gemma 4 models.

## Links

- [DiffusionGemma: 4x faster text generation (Google blog)](https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/)
- [DiffusionGemma — Google DeepMind](https://deepmind.google/models/gemma/diffusiongemma/)
- [DiffusionGemma model overview — Google AI for Developers](https://ai.google.dev/gemma/docs/diffusiongemma)
- [google/diffusiongemma-26B-A4B-it — Hugging Face model card](https://huggingface.co/google/diffusiongemma-26B-A4B-it)
- [DiffusionGemma on NVIDIA NIM](https://build.nvidia.com/google/diffusiongemma-26b-a4b-it/modelcard)
- [DiffusionGemma vLLM recipe](https://recipes.vllm.ai/Google/diffusiongemma-26B-A4B-it)
- [Simon Willison — DiffusionGemma notes](https://simonwillison.net/2026/Jun/10/diffusiongemma/)
- [MarkTechPost — Google AI releases DiffusionGemma](https://www.marktechpost.com/2026/06/10/google-ai-releases-diffusiongemma-a-26b-moe-open-model-using-text-diffusion-for-up-to-4x-faster-generation/)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/diffusion-gemma/