AIDiffusionGemma-Jev gets vLLM support and one-command Cloud Run deployment
vLLM says the model provides confidence with each answer to yes/no, multiple-choice or scored questions. Google Gemma puts Cloud Run costs at roughly $3 an hour, dropping to $0 when idle.
TLDR
vLLM announced support for DiffusionGemma-Jev, saying it provides confidence with answers to yes/no, multiple-choice and scored questions. It describes starting with a response template, leaving only the answer slots noisy and reading a probability distribution from each slot in a single denoising step.
Google Gemma shared code and instructions for launching a Jev API-compatible endpoint on Google Cloud Run with one command, without needing your own GPU. It reported roughly 35–60 ms single-step latency and about 100–123 requests per second at a batch size of 32. It put costs at roughly $3 an hour, dropping to $0 when idle.