Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

djev-run

Deploys DiffusionGemma-Jev behind a TypeSafe-compatible API on a Cloud Run GPU, with snake, dino and tetris demos.

Serviceprototype
View on GitHub ↗

Overview

djev-run serves DiffusionGemma-Jev (djev, from mmastrac/djev) on Google Cloud Run with an NVIDIA RTX PRO 6000 Blackwell GPU, behind a TypeSafe-compatible decision API. The README walks through uploading the NVFP4 model to Cloud Storage and deploying either a pre-built container with all vLLM flags baked in or the upstream vLLM nightly container, following Cloud Run GPU best practices. A Colab notebook is also provided.

Built-in demo apps at /snake, /dino and /tetris are wired to the decision endpoint, inspired by earlier Laya and Jev game demos. It is aimed at people who want a scale-to-zero, self-owned Jev-compatible endpoint on managed GPUs. A second option deploys the upstream vLLM nightly container directly with no custom Dockerfile; the pre-built container is recommended because it bakes in shared-memory staging and all serve flags.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich move should the game agent make next?(inferred)—

How it works

  1. Upload model weights to Cloud Storage
  2. Deploy the container to Cloud Run with a GPU
  3. Send Jev-format requests to the service
  4. Play the bundled demos against the endpoint

Strengths

  • Managed, scale-to-zero GPU hosting
  • TypeSafe-compatible API
  • Ready-made demo games

Caveats

  • Needs a GCP project with RTX PRO 6000 quota in supported regions
  • GPU cost when running

Getting started

Upload nvidia/diffusiongemma-26B-A4B-it-NVFP4 to GCS, then gcloud beta run deploy with ghcr.io/taeold/djev-run:latest.

At a glance

Type
Service
Status
prototype
Language
Python
Runtime
Cloud Run GPU (vLLM)
Has code
yes

Use cases and integrations

How curated lists describe it

  • Serving: deploys DiffusionGemma-Jev behind a TypeSafe-compatible API on a Cloud Run GPU with snake, dino and tetris demos wired to the decision endpoint.

    yibie/awesome-jev

  • Serves DiffusionGemma-Jev behind a compatible API on Cloud Run, with a small game demo on top.

    MrJev/awesome-jev

Frequently asked questions

What model is served?

DiffusionGemma-Jev, a Jev-style decision setup on DiffusionGemma.

Where does it run?

On Google Cloud Run with an RTX PRO 6000 GPU.

What demos are included?

Snake, dino and tetris at /snake, /dino and /tetris.

Related projects

  • laya-mlx

    Independent MLX port of Laya for Apple Silicon: typed decisions in about 13 ms (7 ms multilingual) with no PyTorch or cloud API.

    ServicebetaPython
  • decider

    Open family of System One-style models fine-tuned from Qwen3.5 and Gemma that return typed decision probabilities in one pass.

    ModelbetaPython
  • von

    Apache-2.0 395M ModernBERT System One model serving calibrated Choice, Noul and Score locally as a drop-in /v1/systemone replacement.

    ServicebetaPython, TypeScript
  • openjev

    Gemma 3 4B option scorer that prefills context once and scores all options in one pass, on MLX or PyTorch, with a Doom demo.

    ServiceprototypePython
  • system-one-model

    Gemma 4 12B with its LM head replaced by a small answer head, serving Jev-style noul, choice and score decisions with custom code.

    ModelresearchPython
  • decidealot

    Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.

    MCP serverbetaPython