ACPACP AI Research
⌕ZH
ResourceHeat 85Quality 95

Google DeepMind's DiffusionGemma Open-Source Model

Google DeepMind has released DiffusionGemma, a 26-billion-parameter open-weights model designed for rapid text generation. It utilizes a diffusion technique, enabling significantly faster output compared to traditional autoregressive models, though with a trade-off in quality for certain benchmarks.

AILarge Language ModelsOpen SourceGoogle DeepMindDiffusionGemmaText Generation

Resource Overview

Google DeepMind recently introduced DiffusionGemma, an innovative open-weights model aimed at revolutionizing text generation speed. Released on June 10, 2026, this 26-billion-parameter model represents a significant step in making advanced AI capabilities more accessible to developers and researchers [1].

Key Content

DiffusionGemma distinguishes itself by employing a novel diffusion technique for text generation, moving away from the conventional token-by-token approach. This method, similar to those used in image diffusion models, allows the model to generate up to 256 tokens in parallel per forward pass [1]. This results in impressive speeds, exceeding 1,000 tokens per second on a single Nvidia H100 GPU, making it approximately four to five times faster than comparable autoregressive models [1].

The model's architecture, a mixture-of-experts (MoE) design, ensures efficiency; only 3.8 billion of its 26 billion parameters are active during inference. This design choice facilitates local execution on consumer hardware with 18GB of VRAM when quantized [1].

How to Use

The weights for DiffusionGemma are openly available under an Apache 2.0 license, accessible on popular platforms such as Hugging Face, Kaggle, and Google Cloud's Vertex AI Model Garden [1]. Developers have a wide array of frameworks to integrate and run DiffusionGemma, including Hugging Face Transformers, vLLM, SGLang, MLX, JAX's Hackable Diffusion library, Unsloth for fine-tuning, and Nvidia's NeMo Framework [1].

Notes and Caveats

Google DeepMind has designated DiffusionGemma as an experimental release [1]. While it offers remarkable speed, this comes with a trade-off in output quality; the model performs lower than the standard Gemma 4 model on established benchmarks like MMLU and coding evaluations [1]. For production scenarios where output quality is paramount, Google DeepMind recommends using Gemma 4 [1]. Users should carefully evaluate the balance between generation speed and quality for their specific applications [1].

Sources

  • [1] MLQ.ai, "Google DeepMind Releases DiffusionGemma, a 26B Open-Source Model That Generates Text 4x Faster via Diffusion", published on 2026-06-11.
  • [2] Google DeepMind, "Models", published on 2026-06-01.
  • [3] Google DeepMind, "Gemma", published on 2026-06-01.