Breaking

Google Releases Open-Weight Text Diffusion Model DiffusionGemma, Up to 4x Faster

June 10, 2026 at 12:41 EDT

Foundation Models
Open Source
Infra & Chips

Google DeepMind on June 10 released DiffusionGemma, an experimental open-weight model that generates text in parallel blocks. It claims up to 4x faster generation on dedicated GPUs and is available for immediate download from Hugging Face under the Apache 2.0 license.

June 10, 2026 · Google DeepMind

DiffusionGemma brings image-style diffusion to text — up to 4× faster generation

An open-weight (Apache 2.0) model built on Gemma 4 that generates blocks of ~256 tokens at once and refines them iteratively — instead of one token at a time — running even on consumer GPUs.

4×

faster generation on dedicated GPUs

26B

MoE params (~3.8B active)

18GB

VRAM with quantized build — runs local

256

tokens generated per block, in parallel

Throughput — tokens per second

Cited generation speed by hardware (including overhead).

1,000+

H100

700+

RTX 5090

How it generates text

Generate block

~256 tokens produced at once, in parallel

→

Refine iteratively

Multiple passes with bi-directional attention

→

Self-correct

Real-time fixes, infilling & Markdown

Unlike autoregressive LLMs that emit one token at a time, the diffusion approach drafts a whole block then sharpens it — borrowing the technique used in image generation.

Where it shines

Fast code generation, editing & infilling
Runs locally on ~18GB VRAM
Quantized GGUF builds & RTX optimization
Near-offline, speed-first local experience

Caveats

Quality lags standard Gemma 4
Speed edge may shrink in cloud batch jobs
Experimental — validate before production
No DiffusionGemma-specific benchmark yet

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.