Summary
Google has released DiffusionGemma, an experimental open model that takes a fundamentally different approach to text generation. Instead of the traditional autoregressive method (generating one token at a time), DiffusionGemma uses text diffusion to produce output up to four times faster.
The model is a 26-billion parameter Mixture of Experts (MoE) architecture built on the Gemma 4 backbone, designated DiffusionGemma 26B-A4B. It supports multimodal inputs — text, images, and video — while generating text outputs. The model was released on June 10, 2026, as part of Google’s expanding open-source AI offerings.
DiffusionGemma joins a growing family of Google open models, including the recently released Gemma 4 12B (May 22). Meanwhile, the highly anticipated Gemini 3.5 Pro remains in limited preview for select Vertex customers, with general availability expected later this month.
Source
Commentary
This is a genuinely interesting research direction. Autoregressive generation (the approach used by GPT, Claude, Gemini, and essentially every major LLM) is inherently sequential — each token depends on the previous one, which creates a hard speed ceiling. Text diffusion sidesteps this by refining the entire output simultaneously, similar to how image diffusion models work.
A 4x speed improvement is significant enough to matter for production workloads, especially latency-sensitive applications like real-time chat and coding assistants. The fact that Google is releasing this as an open model suggests they see diffusion-based text generation as a complement to — not a replacement for — their proprietary Gemini line. If the quality holds up under scrutiny, expect other labs to accelerate their own diffusion-for-text research.