DiffusionGemma: 4x Faster Local Text Generation with Google Text Diffusion

---

What is DiffusionGemma: Overview of Text Diffusion Models

Google DeepMind announced DiffusionGemma on June 10, 2026, as an experimental open-source model. It adopts a completely different approach from traditional autoregressive LLMs by using a text diffusion process that enables parallel generation of 256 tokens.

📑Table of Contents

1. What is DiffusionGemma: Overview of Text Diffusion Models
2. Performance Benchmarks and Hardware Requirements
3. Differences from Previous Gemma/Gemini and Positioning
4. Use Cases and Developer Practicality
5. Limitations and Caveats
6. Getting Started with DiffusionGemma and the Ecosystem
7. Frequently Asked Questions (FAQ)
8. Summary

With 26B total parameters but only 3.8B active during inference thanks to its MoE structure, it is released under the Apache 2.0 license. The official announcement is available at https://deepmind.google/models/gemma/diffusiongemma/.

---

Performance Benchmarks and Hardware Requirements

On NVIDIA H100 it achieves over 1,000 tokens/sec, and on RTX 5090 it reaches 700+ tokens/sec — up to 4-5x faster than previous autoregressive models.

Item	Value	Notes
Total Parameters	26B	MoE
Active	3.8B	Inference
Max Speed	1000+ tok/s	H100
VRAM Requirement	~18GB	Quantized
Parallel Generation	256 tokens	1 pass

Source: https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/ , https://ai.google.dev/gemma/docs/diffusiongemma (as of June 2026)

---

Differences from Previous Gemma/Gemini and Positioning

Although based on Gemma 4, the generation method is fundamentally different, using bidirectional attention to consider the entire context simultaneously. Its self-correction capability allows iterative refinement of entire blocks, excelling at format closing and error correction.

It is optimized for local low-concurrency workloads and is not suited for high-QPS cloud use. While quality is slightly lower than Gemma 4, it excels in interactive real-time applications.

---

Use Cases and Developer Practicality

Ideal for inline editing, code infilling, mathematical graphs, and Sudoku-like dependency tasks. Parallel generation provides advantages in non-linear tasks. It is effective for local GPU operation and building high-speed AI assistants.

Recommended settings include max 48 denoising steps, temperature schedule 0.8→0.4, and adaptive early stopping (entropy threshold 0.005).

---

Limitations and Caveats

Efficiency drops with short generations, and a single token error can invalidate the entire output. For large-scale cloud batch processing, autoregressive models remain advantageous. As an experimental model, Gemma 4 is recommended for production use.

---

Getting Started with DiffusionGemma and the Ecosystem

Weights can be obtained from Hugging Face, Kaggle, or Vertex AI Model Garden. Inference supports Hugging Face Transformers, vLLM, and MLX, with llama.cpp support coming soon. For fine-tuning, Unsloth, Hackable Diffusion (JAX), and NVIDIA NeMo are available.

---

Frequently Asked Questions (FAQ)

Q: How is DiffusionGemma different from Gemma 4?

A: The biggest difference is the adoption of a diffusion process instead of autoregressive generation. Parallel generation and self-correction are its strengths.

Q: Does it really become 4x faster on local GPUs? What are the hardware requirements?

A: Up to 4-5x speed improvement has been confirmed on H100 and RTX 5090. Quantized models require approximately 18GB VRAM.

Q: What specific advantages does it offer for code completion and inline editing?

A: Parallel generation excels at non-linear tasks by considering future tokens, delivering high accuracy in inline editing and code completion.

Q: Where can I download the quantized models?

A: Available on Hugging Face and Kaggle. Vertex AI Model Garden can also be used.

Q: Why is it more suited for local rather than cloud inference?

A: It is optimized for low-concurrency workloads and excels in interactive real-time applications. It is not suited for large batches.

Q: Is migration from existing Gemma models easy?

A: Some changes to the inference engine are required, but migration is relatively straightforward with Hugging Face Transformers.

Q: Are there fine-tuning examples or guides?

A: Guides using Unsloth and NVIDIA NeMo have been published. Hackable Diffusion based on JAX is also available.

Q: How does it perform with short prompts?

A: Performance tends to degrade with short generations. It shines with longer contexts.

---

Related articles:

• Google Gemini 3.1 Pro の Gemini-SQL2 が BIRD Text-to-SQL で 80.04% SOTA 達成
• Google、Gemini 3.5 Live Translate発表 — 70言語以上対応のリアルタイム音声翻訳
• GLM-5.2 リリース — 1Mトークン context + 長時間エージェント/コーディング特化（MITオープンウェイト）

Summary

DiffusionGemma expands the possibilities of local AI through parallel generation via text diffusion. It is a strong option for developers seeking local GPU operation or high-speed interactive workflows. Start by obtaining the weights from Hugging Face and give it a try.

Related new article:

• Google Gemini Managed Agents API GA — Execute Fully-Managed Agents via Single API Call in Secure Sandbox – This published update adds current operational context for DiffusionGemma: 4x Faster Local Text Generation with Google Text Diffusion.
• Gemini 3.5 Flash Adds Native Computer Use — Cross-Platform Agent Development Made Easier – This published update adds current operational context for DiffusionGemma: 4x Faster Local Text Generation with Google Text Diffusion.
• How to Use Gemini Study Notebooks for Personalized Learning and Practice Quizzes – This published update adds current operational context for DiffusionGemma: 4x Faster Local Text Generation with Google Text Diffusion.
• agents-cli v1.0.0: Official Google CLI to Scaffold, Evaluate & Deploy Production AI Agents (Claude Code / Cursor Compatible) – This published update adds current operational context for DiffusionGemma: 4x Faster Local Text Generation with Google Text Diffusion.