Breaking

Google Gemma Spotlights Jev-Like DiffusionGemma Mode for Parallel Decisions

September 18, 2026 at 18:06 EDT

Foundation Models
AI Agents
Software Dev & Coding

Google Gemma has highlighted a Jev-like use of DiffusionGemma that turns Google's open text-diffusion model into a rapid structured-decision engine. The underlying work is an open vLLM pull request, not a finished product release, and it shows how a non-autoregressive model can score bounded choices without writing a normal response token by token.

Sep. 18, 2026 - DiffusionGemma

Parallel decisions, not prose

A Jev-like vLLM patch turns DiffusionGemma's diffusion canvas into fixed answer slots, reading choice probabilities instead of generating a normal response.

claimed peak generation speedup

256

tokens per diffusion canvas

162/s

reported decisions on DGX Spark

0.12s

reported single-read latency

The speed claim, drawn to scale

autoregressive baseline

DiffusionGemma claim

The claimed peak gain comes from processing a block of tokens in parallel instead of one token at a time.

How the Jev-like read works

Seed schema

Place fixed answer slots in the canvas.

Denoise once

Run a read-only diffusion pass.

Read scores

Use log probabilities and entropy.

Return choices

Emit typed labels with confidence policy.

Why developers care

Moderation, routing, support triage, evals, and game agents can use fast bounded decisions instead of full text generation.

What still needs proof

Calibration, question independence, single-token labels, and production reliability remain open engineering questions.

Continue reading

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