A Cloud Run recipe serves DiffusionGemma-Jev at 35 to 60 ms

Google’s Gemma account posted on September 22 that DiffusionGemma-Jev can be deployed to Cloud Run with one command. The post puts single-step latency at about 35 to 60 ms, batch throughput at 32-wide concurrency at about 100 to 123 requests a second, and the bill at roughly $3 an hour, dropping to $0 when the service is idle. The next post credits Matt Mastracci and Daniel Lee, and quotes Lee’s September 20 note. That note links taeold/djev-run. GitHub lists no license on the repository.

The README serves ghcr.io/taeold/djev-run:latest, which it says is built from mmastrac/djev-spark. The deploy command asks for one nvidia-rtx-pro-6000, 20 vCPU, 80 GiB of memory, concurrency 32, and a maximum of one instance. Minimum instances is 0. The regions it names for that GPU are us-central1, europe-west4, asia-southeast1, and asia-south2. Weights are nvidia/diffusiongemma-26B-A4B-it-NVFP4. The page describes them as 17.5 GB, copied into /dev/shm while Python imports torch, at about 1.05 GiB a second from Cloud Storage. Cold start from zero instances is about 47.5 seconds. The page says an earlier baseline was 4 minutes 5 seconds, and that baking the weights into the image took 80 to 115 seconds.

The latency line matches the post, with one extra setting: about 35 to 60 ms when steps=1 and samples=1, and about 160 ms when samples is auto. Batch throughput is about 100 to 123 requests a second at concurrency 32. The price on the page is $3.19 an hour while the GPU is active. The post’s figure is “roughly $3.”

The server implements POST /v1/systemone. A sample uses @ai-sdk/typesafe-ai and points baseURL at the Cloud Run /v1 path, then calls evaluationModel('jev-latest'). snake.html is a single file that calls the same route from the browser, at about 15 moves a second. The README prints no accuracy table and no JevBench row.

Maisa’s hosted djev is a different service. JevBench lists that API at 74.3, with hard-tier accuracy 69.5% and a p50 of 0.24 seconds. The local stack at Davipar/djev-dev speaks POST /v1/request. This recipe speaks /v1/systemone on a GPU billed by the hour. The 35 to 60 ms cell is the samples=1 note on an RTX PRO 6000, not the hosted bench timing. We did not deploy it.