Skip navigation
Deploying DiffusionGemma-Jev (djev) just got a lot easier.
0:00 / 1:03

Deploying DiffusionGemma-Jev (djev) just got a lot easier.

Google Gemma

@googlegemma111K followers

Visit on X
6,950 likesDownloadOriginal post

647,721 views

Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command. Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec. It's a straightforward way to experiment without needing your own GPU. Runs at roughly $3/hr and drops to $0 when idle. Get the code and instructions here:

31 Comments

Comments from the original post on X (Twitter)

  • Google Gemma's profile picture

    @googlegemma

    Credits to @mmastrac and @dylayed. Original post:

    22 likes

  • Jonathan Ouyang's profile picture

    @jonsouyang

    @bodonoghue85 Fastest bookmark of my life

    10 likes

  • Viren Khandal's profile picture

    @_VirenKhandal

    jev on blackwell in the big 2026 🥀 what are we doing

    9 likes

  • ORC KILLER's profile picture

    @ORC_Killer

    but why ? jev is cheaper than $3/h and way better at it

    5 likes

  • Everlier's profile picture

    @Everlier

    Infinite Jev unlocked

    4 likes

  • hey's profile picture

    @duummbass

    Put some pressure on Gemini team so you guys can distill for Gemma 5 soon 😌

    3 likes

  • iMuffin's profile picture

    @iMuffined

    this is exactly, what i asked for. thanks!

    1 like

  • Vincent Myshinkye's profile picture

    @MyshinskyYA

    This is the kind of deployment UX that makes experimentation actually accessible. Spin it up when you need it, pay while it runs, and let it disappear when you’re done. 🔥

    1 like

  • Ibrahima Sarr's profile picture

    @ibs4lif

    When I was seeing Jev explanation I was asking myself, isn't what DiffusionGemma already doing. Is it the similar, is it that Jev just has more hype?

    1 like

  • Red Wizard 🔮 (e/acc)'s profile picture

    @RightHonorableB

    Coming to the Gemini API/AI Studio?

    1 like

  • Sadnan's profile picture

    @ntf__sadnan

    lmao, didnt see this coming

    1 like

  • The AI Therapist's profile picture

    @TheAIShrink

    Single-ste latency is the real flex. 35ms means you can feel the stream before your coffee cools, smooth workflow vibes that actually respect a mom’s attention span. Efficient code looks good on everyone. 😊

    1 like

  • TensorQuay's profile picture

    @tensorquay

    One practical limit in the repo: the Cloud Run recipe is set to one GPU instance and 32 concurrent requests. The 100–123 req/s figure is a warm batch result; it says little about the wait a new caller sees when that GPU is busy.

    1 like

  • pysolin's profile picture

    @pysolin

    google's like you want ML, we got ML

  • 𓂀♎︎𓂀's profile picture

    @formlesstrades

    Oh ok Google.

  • Bruno B. Sant' Anna's profile picture

    @brunob_santanna

    Interesting! How does this compare to running similar models on other serverless platforms like AWS Lambda or Azure Functions in terms of cost and cold start times?

  • Kyle Kelley's profile picture

    @KyleRayKelley

    @mmastrac you mad lad

  • John Rood's profile picture

    @johnroodepic

    Don't add --cpu-boost by reflex here. The repo reports 73-75s cold starts with it versus 47.5s without it on the 20-vCPU GPU setup.

  • Jhon Dennis's profile picture

    @JhonDennisxiud

    一条命令就能把实验跑起来 这种少折腾的升级最有用 普通人不用先学一堆黑话

  • 刘朝 Zhao Liu's profile picture

    @liuzhao_666

    The interesting part is the deployment contract: Jev-compatible API, scale-to-zero, and a single Cloud Run command. The next useful artifact is a cold-start and tail-latency trace under real bursty traffic, because “$0 when idle” matters only when wake-up cost stays predictable.