Breaking

Google Releases Open Diffusion Model DiffusionGemma

June 10, 2026 at 12:07 EDT

Foundation Models
Open Source
Media Generation

Google DeepMind on June 10, 2026 released DiffusionGemma, an experimental open diffusion model that generates text in parallel, under an Apache 2.0 license. Moving away from the conventional token-by-token approach, it generates entire blocks of text simultaneously, claiming up to 4x faster generation.

June 10, 2026 · Google DeepMind

DiffusionGemma: text that prints, not types

An experimental open model built on Gemma 4 (Apache 2.0) generates whole blocks of text at once — refining a noisy canvas like a printing press instead of predicting one token at a time, aiming for up to 4× faster generation.

4×

faster vs. AR + MTP drafter on dedicated GPUs

~256

tokens handled at once per block

Apache 2.0

free commercial use, modification, redistribution

Measured generation speed

Tokens per second — taller column means faster output.

700+

RTX 5090
tokens/s

1000+

H100
tokens/s

Two ways to generate text

Autoregressive (AR)

token → token → token
predicts the next single token, one step at a time

DiffusionGemma

noisy canvas → refine → refine → finished block
a whole block emerges at once, like a printing press

Why it matters

Real-time formatting of complex Markdown
Self-correction mid-generation
Pushes the AR intelligence–latency frontier
Open weights, RTX PRO & DGX optimized

The caveats

Speed may come at a quality tradeoff
Apple Silicon: memory bandwidth bottleneck limits gains
Best on dedicated NVIDIA RTX / DGX GPUs
Fresh release — broad usage feedback still limited

Continue reading

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