User repository on GitHub
DiffusionGemma NVFP4 structured decisions on a DGX Spark: container recipe
- Forks
- 16
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Models & Reimplementations
Djev Spark
A Docker Compose image that serves DiffusionGemma 26B-A4B on a DGX Spark behind Jev's systemone endpoint, adding seven not-yet-upstream vLLM perf branches and dependency-keyed multi-stage question reads.
- Stars
- 221
- Language
- HTML
- Published by
- Community
- Author
- mmastrac
- Last verified
- Oct 2026
About this project
DiffusionGemma 26B-A4B (NVFP4) on a DGX Spark, serving structured decisions through Jev's API. One container runs vLLM on port 8010 and the structured decision server on port 8011 in front of it. The server is mmastrac/djev; it implements Jev's POST /v1/systemone API and passes ordinary chat through to vLLM. Structured reads are upstream in vLLM (vllm-project/vllm#57250). This image adds seven perf branches that are not upstream yet; see docs/vllm-stack.md.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- vLLM runs on port 8010 with a structured decision server on 8011 implementing the systemone API and passing ordinary chat straight through.
- The image carries seven perf branches not yet upstream, and the build aborts if a branch's change is missing or a patched module fails to import.
- Questions read together share one canvas and cannot see each other, while dependency, gating and isolation keys split them into staged or separate reads.
- Needs a GB10 box with a CUDA 13 driver plus about 25 GB for the image and 18 GB for the checkpoint; other hardware is untested.
Watch out
This targets DGX Spark or other GB10 hardware, and the README says other boxes may work without saying what would need to change. It rides a nightly vLLM base plus a fork overlay of seven perf branches, so behaviour tracks upstream movement and the bundled status script exists to tell you when a branch has been merged. Image input is an extension rather than part of Jev's API, sequential reads reject it, and the diagnostics field is explicitly server-specific and subject to change. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
For the core surface, yes. The server takes Jev's request body and returns Jev's answer shapes on the systemone endpoint, and the model field is accepted purely for compatibility and then ignored, so an existing client pointed at this endpoint should not need edits. It requires no API key unless you set one, and ordinary chat requests pass through to the local vLLM server behind it.
The README presents no quality comparison, and that is the real gap. What is documented is the serving shape, the seven perf branches layered on a nightly vLLM build, an optional seed that makes a repeated request give the same answer, and how usage is counted, with output tokens counted as canvas rows read plus any thought tokens. Nothing here benchmarks answer quality.
This server's invention, and the README labels them extensions beside state and questions: sampling count with automatic re-reads on high entropy, a think budget, shared-prefix instructions, chunking controls, sequential mode, a subset selector and extra denoise steps. Dependency, gating and isolation keys are likewise additions, so a client written against stock Jev will ignore them and a client written for this server will not port.
Because it is a full serving stack rather than a thin adapter. The image bundles the nightly vLLM base, the fork overlay, the structured server and helper scripts, and the checkpoint is the quantised DiffusionGemma weights. Memory beyond that is the weights plus the KV pool you size yourself and a start-up transient that scales with max sequences and canvas, which a variable exists to budget.
One build-time patch stops the build if the base vLLM is not the stack's commit, and another stops it if any branch's change is missing from the installed vLLM, if a changed module fails to import, or if the server lacks a flag the entrypoint passes. A status script reports per branch whether the fork moved and whether upstream has merged the PR yet.
seed seeds the noise draws, so the same request with the same seed gives the same answer, defaulting to 42. Other top-level extensions include a samples count with automatic re-reads on high entropy, a think budget, shared-prefix instructions, chunking controls, sequential mode, an ask subset selector, and extra denoise steps.
As a server extension. Send multipart form data with the JSON body in a request part and each image as a file part, or an images array of data URLs or content-type and base64 objects, and the server puts them ahead of the state in the prompt. sequential mode needs a text-only state and returns 422 with images.
More Jev resources like this
1.7kGitHub stars
Open-weight 4B and 9B models from TokenRhythm: NeoHorse-1 for tool use, coding, and instruction following, plus NeoHorse-Jev-4B for prefill-only Choice, Noul, and Score decisions.
Jev Models & Reimplementations#community#jevbest
Community
TokenRhythm/NeoHorsePython
TokenRhythm/NeoHorsePython1.5kGitHub stars
A 0.8B decision model that answers first and escalates only when unsure, plus nine LoRA adapters of roughly 41 MB each for guard, triage, tool choice, spam, and similar classification work.
Jev Models & Reimplementations#community#jevbest
Community
firelex/jeffPython
firelex/jeffPython2kGitHub stars
DeepOpen is a non-autoregressive decision engine built on Laya that classifies into developer-defined types in one forward pass, with a language router picking between English, multilingual, and typed-decision checkpoints.
Jev Models & Reimplementations#community#jevbest
Community
deepopen-com/deepopenPython
deepopen-com/deepopenPython715GitHub stars
Decision models that answer several choice, score, and noul questions over a state with probabilities in one forward pass, with training, two inference backends, calibration tooling, and the benchmark suite in the repo.
Jev Models & Reimplementations#community#jevbest
Community
InternLM/Intern-DecisionPython
InternLM/Intern-DecisionPython402GitHub stars
Decision models that return typed probabilities for candidate labels from a state, instructions, and options, shipped as LoRA adapters plus a trained decision head over pinned Qwen base weights.
Jev Models & Reimplementations#community#jevbest
Community
Zefan-Cai/Open-JevPython
Zefan-Cai/Open-JevPython363GitHub stars
Open-weight 2B, 4B and 9B decision models that extend Jev's typed question format to photos, so an image plus a record can be checked against options you supply, with a probability on each and an explicit can't-tell. Runs on your own hardware.
Jev Models & Reimplementations#community#jevbest
Community
mohit67890/imajevPython
mohit67890/imajevPythonFrom the Jev guides
Original write-ups on what this entry is for.
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.
Jev Models & Reimplementations#decision-model#benchmarks#pricing
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.