Breaking

Google Releases DiffusionGemma, an Open Model Using Text Diffusion

June 10, 2026 at 12:44 EDT

Foundation Models
Open Source
Media Generation

On June 10, 2026, Google released DiffusionGemma, an open-weight model under the Apache 2.0 license that generates blocks of text in parallel using "discrete text diffusion" rather than predicting tokens one by one. It claims up to 4x faster generation on dedicated GPUs/TPUs, while its quality benchmarks trail the underlying Gemma 4.

June 10, 2026 · Google

DiffusionGemma: an open model that writes text in parallel, not word-by-word

Google's new Apache 2.0 open-weight model swaps next-token prediction for image-style diffusion — generating 256-token blocks at once and refining them. The claim: up to 4× faster output.

4×

faster output vs. sequential generation

26B

MoE params · 3.8B active at inference

18GB

VRAM when quantized — fits one GPU

Inference speed by GPU

tokens generated per second — taller column = faster

1,000+

H100

data-center

700+

RTX 5090

consumer

How it generates: blocks, not one word at a time

Draft 256-token block in parallel

→

Iteratively refine & self-correct

→

Output with real-time Markdown formatting

Where it shines

Low-latency local / edge inference
Code infilling & in-line editing
Constrained tasks — 80% success on a fine-tuned Sudoku run
Single-GPU workloads (RTX 4090 / 5090)

The trade-off

Output quality lags standard Gemma 4
Best for single local GPU, not batched cloud use
Detailed real-world reports still limited
Speed-vs-quality balance remains a challenge

Seen as an early sign of a shift from autoregressive to diffusion-based text generation — released under Apache 2.0 for commercial use, modification and redistribution.

Continue reading

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