Breaking

vLLM Merges DiffusionGemma Structured Reads for Jev-Style Decisions

September 29, 2026 at 14:52 EDT

Foundation Models
AI Agents
Open Source

vLLM has merged structured-read support that lets DiffusionGemma operate as a Jev-style decision engine, returning bounded answers and probability distributions instead of prose. The change is significant because it moves a fast decision-model pattern from hosted early-access APIs toward self-hosted open weights, while still leaving the endpoint in example-server territory.

Sep. 2026 · vLLM + DiffusionGemma

Structured decisions, not prose

A merged vLLM change seeds DiffusionGemma's canvas with an answer template, then reads probabilities from fixed slots after one denoising step.

denoising step for a structured read

requests per second in the 32-way DGX Spark test

162

decisions per second at three questions per request

token decision canvas in the nightly recipe

Throughput scales with concurrent reads

8.7 req/s1-way

54 req/s32-way

The visual shows the same unit side by side: requests per second rose from the single-request run to the 32-concurrent run.

How the decision read works

1 Seed template

Fixed text is placed in the canvas.

2 Leave slots noisy

Only answer positions remain open.

3 Denoise once

The model reads the whole canvas.

4 Return probabilities

Logprobs become decisions and uncertainty.

Why builders like it

Open weights can support local routing, moderation, scoring, and guardrails on infrastructure teams control.

What remains early

The endpoint is an example server, labels must be single tokens, and thresholds need real workload validation.

Continue reading

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