Google DeepMind released DiffusionGemma — an open 26B MoE model with 4x faster generation

Google DeepMind released DiffusionGemma — an experimental open-source language model with 26 billion parameters that uses text diffusion instead of the conventional autoregressive generation. On GPU, it runs up to four times faster than standard approaches.

What is text diffusion

Most modern language models generate text token by token from left to right — this is how GPT-4, Gemini, Llama, and virtually all major LLMs work. This is reliable and well-studied, but this approach has a fundamental limitation: inference speed scales linearly with answer length. The longer the text, the longer the wait, the higher the GPU costs.

DiffusionGemma works differently. The model starts with a noisy or masked output and iteratively refines it until coherent text emerges — analogous to how diffusion models like Stable Diffusion generate images. The key difference from autoregression is parallelism: instead of strict sequentiality, the diffusion decoder can work across the entire context at once. This is what delivers the multiplicative speed gain on modern GPUs.

Text diffusion research has been underway for several years, but large-scale open models of this class have been scarce. DiffusionGemma is one of the first serious public experiments of this magnitude from a major lab, and deserves attention for that reason alone.

Architecture: 26B with MoE

DiffusionGemma is built on a Mixture of Experts (MoE) architecture. Unlike "dense" models where all parameters are engaged on every request, MoE activates only a subset of expert blocks — depending on input data. This allows for a large number of parameters at relatively low computational cost during inference.

Key model characteristics:

• 26B parameters total (MoE architecture)
• Only a portion of parameters activated during inference
• Text diffusion instead of autoregression
• Up to 4× speedup in generation on GPU
• Open access for researchers
• Experimental status — not a product release

Combining MoE and diffusion is a non-trivial architectural bet. MoE reduces the burden from the number of engaged parameters, diffusion reduces it from the number of generation steps. In theory, both improvements work together.

Why this changes the picture

Inference speed is one of the major practical challenges for large language models. For end users, slow responses are frustrating. For inference providers, it means direct GPU-time costs that directly impact service margins. Current solutions — quantization, speculative decoding, optimized kernels — yield speedups around 1.5–2×. DiffusionGemma claims 4×, through a fundamentally different generation mechanism. If this holds up in real conditions, we're talking about a paradigm shift, not an optimization.

Google DeepMind is releasing the model in open access as a research artifact. This gives the academic community the opportunity to study a 26B-scale diffusion text decoder. The question of a product version based on this architecture remains open.

What this means

DiffusionGemma signals that autoregression is ceasing to be the only viable paradigm for language modeling. If the diffusion approach scales without quality degradation, the response speed of AI tools could increase manyfold — without proportional growth in infrastructure costs. The community's investigation and benchmarking of the model in coming months is worth monitoring.