DiffusionGemma runs as a parallel decision model, faster than autoregressive generation

AISummary

Google Gemma's account says DiffusionGemma, running in a Jev-style decision setup, denoises an open canvas in one step rather than generating tokens sequentially, taking about 0.2 seconds on a DGX Spark.

It says full bidirectional attention lets every option attend to the full context at once, and that the model inherits Gemma 4's spatial vision capabilities for visual and text decisions.

Post on XView on X

Google Gemma

@googlegemma

"DiffusionGemma as Jev" showcases the power of non-autoregressive architectures.

While Jev demonstrates the value of rapid decision models, running DiffusionGemma in this paradigm leverages canvas diffusion to evaluate structured choices in a single parallel pass:

⚡ ️Massive Parallelism: Denoises across an open canvas in a single step instead of sequential autoregressive token generation (~0.2s on a DGX spark).

🧠 Full Bidirectional Attention: Allows every option to attend to the full context concurrently, yielding well-calibrated decision distributions.

👁️ Multimodal Grounding: Inherits Gemma 4's spatial vision capabilities for complex visual and text decisions.

Image from @mmastrac's post — Matt Mastracci@mmastrac

Source: Google Gemma · x.comPublished Sep 18, 6:04 PM ET · added here Oct 9, 8:08 AM ET

DiffusionGemma runs as a parallel decision model, faster than autoregressive generation

Topics