Breaking

Google releases DiffusionGemma, an open LLM built on diffusion

June 10, 2026 at 12:07 EDT

Foundation Models
Open Source

Google DeepMind on June 10, 2026 released DiffusionGemma, an experimental open-weight model under the Apache 2.0 license that applies diffusion modeling to text generation. It claims up to 4x faster output on dedicated GPUs and is available immediately on Hugging Face.

June 10 · Google DeepMind

DiffusionGemma: Text Generation That Refines Instead of Predicting

An experimental open-weight model built on Gemma 4 generates whole blocks of 256 tokens at once and iteratively denoises them — claiming up to 4× faster output on dedicated GPUs versus conventional one-token-at-a-time methods.

4×

Faster than AR + Multi-Token Prediction

256

Tokens generated per block, then refined

Apache 2.0

Open weights — commercial use & redistribution

Throughput on dedicated GPUs

Tokens generated per second (reported figures)

700+

RTX 5090

tokens/s

1000+

H100

tokens/s

Two ways to generate text

Conventional (AR)

token→ token→ token

One token at a time — sequential, so latency stays high.

DiffusionGemma

256-token block→ denoise / refine

Whole block at once, then iteratively refined — enabling self-correction during generation.

Promise

High parallelism, major GPU acceleration
Real-time Markdown formatting
Improved long-text consistency
Pushes the latency–quality Pareto front

Open questions

Possible quality drop vs Gemma 4
Extra compute from denoising steps
Unclear practicality on smaller GPUs
No benchmarks, parameter count or training data disclosed yet

Benefits are maximized on high-performance NVIDIA GPUs.

Performance on CPUs and low-spec setups remains undetailed, and concrete reports for code, writing and summarization are still scarce — community reproduction is awaited.

Continue reading

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