---
title: "DiffusionGemma: The Developer Guide"
description: "DiffusionGemma is an experimental text-generation model and developer guide covering serving, customization, and non-autoregressive generation. Built on the Gemma 4 backbone, its 26B Mixture of Experts architecture activates 3.8B parameters during inference, refines 256-token canvases in parallel, and shifts the bottleneck from memory bandwidth to compute. Block autoregressive diffusion commits denoised 256-token blocks to the KV cache, combining bidirectional denoising with causal prefill for longer sequences; vLLM integration provides an OpenAI-compatible local server. A Sudoku fine-tuning example using Hackable Diffusion reports improvement from roughly 0% base-model success to 80%, while SFT reduces the demonstrated solve from 48 to 12 steps through adaptive stopping. The guide also identifies bidirectional attention and re-noising as mechanisms for global constraint handling and self-correction."
---

# DiffusionGemma: The Developer Guide

[Google](https://yomu.fyi/company/google) · Ian Ballantyne, Omar Sanseviero · Jun 10, 2026

**Type:** Tutorial

## Summary

DiffusionGemma is an experimental text-generation model and developer guide covering serving, customization, and non-autoregressive generation. Built on the Gemma 4 backbone, its 26B Mixture of Experts architecture activates 3.8B parameters during inference, refines 256-token canvases in parallel, and shifts the bottleneck from memory bandwidth to compute. Block autoregressive diffusion commits denoised 256-token blocks to the KV cache, combining bidirectional denoising with causal prefill for longer sequences; vLLM integration provides an OpenAI-compatible local server. A Sudoku fine-tuning example using Hackable Diffusion reports improvement from roughly 0% base-model success to 80%, while SFT reduces the demonstrated solve from 48 to 12 steps through adaptive stopping. The guide also identifies bidirectional attention and re-noising as mechanisms for global constraint handling and self-correction.

## Context

Traditional autoregressive models repeatedly load weights from memory and generate one token at a time, making memory bandwidth a primary GPU-serving bottleneck. Their left-to-right generation also makes strict, multivariable constrained tasks such as Sudoku difficult because future positions cannot be evaluated or revised.

## Approach / What changed

DiffusionGemma uses Uniform State Diffusion to iteratively refine 256-token canvases in parallel, then applies block autoregressive diffusion for longer sequences by committing denoised blocks to the KV cache. The guide covers bidirectional denoising, causal prefill, vLLM serving, and Sudoku fine-tuning with Hackable Diffusion and SFT.

## Takeaways

- DiffusionGemma's 26B MoE architecture activates 3.8B parameters during inference, enabling quantized deployment within 18 GB of VRAM.
- The model combines bidirectional attention during denoising with causal incremental prefill, allowing canvas-wide context, re-noising for self-correction, and sequential handling of longer outputs.
- On Sudoku, the base model reports approximately 0% success, while a simple JAX SFT recipe raises correctness to 80%; the demonstrated fine-tuned run finishes in 12 steps versus 48 for the base model.

**Tags:** [Generative AI](https://yomu.fyi/topic/generative-ai), [GPU](https://yomu.fyi/topic/gpu), [LLMs](https://yomu.fyi/topic/llm), [Model Serving](https://yomu.fyi/topic/model-serving), [vLLM](https://yomu.fyi/topic/vllm)

- Source: [Google](https://developers.googleblog.com/diffusiongemma-the-developer-guide)
- Source URL: https://developers.googleblog.com/diffusiongemma-the-developer-guide
- Ingested by Yomu: 2026-09-09T14:29:48.940Z

[Read original post](https://developers.googleblog.com/diffusiongemma-the-developer-guide)