DiffusionGemma

// 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

This model's scores

1. MMLU-Pro77.6%
2. GPQA Diamond73.2%
3. AIME 2026 (no tools)69.1%
4. LiveCodeBench v669.1%
5. MATH-Vision70.5%
6. MMMU Pro54.3%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

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.

Pricing source ↗

// 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

// Gemini Diffusion / DiffusionGemma — every version

The full lineage of the Gemini Diffusion / DiffusionGemma line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.