User repository on GitHub
A Jev-like decision engine powered by DiffusionGemma and SGLang. Jev with eyes.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Diffusion Jev Sglang
An independent typed-decision server built on Google's DiffusionGemma and SGLang: it reads image pixels as well as text, returns choice, noul and score answers, and ships a drawing playground and local evaluation artifacts.
- Stars
- 10
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- Hangzhi
- Last verified
- Oct 2026
About this project
A diffusion model that guesses your doodles, recognizes flowers, and picks an emoji for your message. Built with DiffusionGemma and SGLang. Try the playground(Cold start takes minutes)· Model · Results · Setup guide. One model · Three demos · No fine-tuning. A wonky sketch goes in. Cat comes out, with scores for all 16 choices. Draw with a mouse or a finger, upload your own image, or try one of 384 Google Quick, Draw! sketches. Click the GIF to open the drawing pad.
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- Three demos off one model: 16-class doodle recognition, five-way flower classification, and text that yields an emoji, a yes/no answer and a score.
- Options are mapped to answer letters and the adapter reads letter scores from the last active denoising step; each question is its own model request.
- On 231 public JevBench tasks the local DiffusionGemma 26B A4B run gets 196 correct, against 152 for local LLaDA2.1-mini.
- Separate tests report 98 of 100 flowers, 42 of 50 Emojify and 311 of 1000 TweetEval, where 105 output-format failures are counted as wrong.
Watch out
This is an independent project with a Jev-like API, not TypeSafe's Jev, and it runs Google's published DiffusionGemma weights; the compared Jev 1.13.0 and LLaDA2.1-mini interfaces are text-only, so the image demos are not part of the benchmark. The scores compare the supplied options and are explicitly not calibrated probabilities of being correct. The accuracy and latency comparison is not apples to apples: local runs used an A100 while the official Jev numbers come from an older published API run, the README says the timings do not establish which model is faster, and no fresh hosted Jev comparison has been completed. Doodle Detective has no held-out accuracy result. Reproducing it means one A100 80GB, Linux, Python 3.12, roughly 50 GB of weights on disk and about 72 GiB of GPU memory reserved, against a pinned experimental SGLang revision carrying this repo's readout patch. The public playground shares an allowance of 100 prediction jobs per day and 1000 per month, starts an A100 on demand and can take minutes on a cold start. Project code is MIT; weights and datasets carry separate licenses. 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.
Treat it as indicative, not a head-to-head. The local DiffusionGemma run scored 196 of 231 JevBench tasks and local LLaDA2.1-mini scored 152, but the official Jev 1.13.0 figure of 200 comes from an older published API run rather than a fresh one. Different hardware and network paths mean the latency columns do not establish which model is faster either, and the README says so directly. The public playground adds queueing, startup and network time on top of that.
Images are the point of the model choice. DiffusionGemma reads the image pixels alongside your question, which is how a daisy photo resolves to daisy out of dandelion, rose, sunflower and tulip, and how a sketch lands on cat out of the 16 Quick Draw classes. What is text-only is the comparison: the Jev 1.13.0 and LLaDA2.1-mini interfaces used in the benchmark take no pixels, so the image results are separate tests rather than part of the JevBench table.
The tested setup is a single NVIDIA A100 80GB on Linux with Python 3.12, about 50 GB of disk for weights, and an engine that reserves roughly 72 GiB of GPU memory. You run uv sync --locked, then a launch script that needs three paths filled in: the model, your SGLang source tree and the engine interpreter. The web app is included, the local decision endpoint is POST /v1/systemone, and there is an SSH forwarding guide if the GPU is on another machine.
Not as probabilities. The README states the scores compare the supplied options and are not calibrated probabilities of being correct, and that a high score can still be wrong. It also counts 105 TweetEval output-format failures as wrong answers, which is a useful reminder that a correct-looking distribution can still produce unusable output on a subset of tasks.
Your options are mapped to letters, A for cat and B for clock and so on. DiffusionGemma runs its denoising steps, and the adapter reads the answer-letter scores from the last active denoising step before mapping them back to your labels. Each question is its own model request, which SGLang can batch. It reuses the published model rather than training a new decision head.
The playground is hosted on Vercel with Modal starting a single A100 only when you request a prediction and a 60-second idle window, so a cold start can take minutes. The public allowance is shared across all visitors at 100 prediction jobs per day and 1,000 per month, and a separate monthly compute budget can pause predictions earlier. Drawing and browsing do not start the GPU.
A sketch must resolve to one of 16 Google Quick, Draw! classes: airplane, apple, bicycle, cat, clock, fish, pizza, umbrella, dog, car, house, tree, sun, star, cup, and sailboat. You can draw, upload, or try one of 384 provided sketches. Doodle Detective has no held-out accuracy result, which is the one number the README refuses to give.
More Jev resources like this
48GitHub stars
Self-contained MuJoCo reproduction that uses two sequential Jev Choice calls per update for manipulation intent and axis plus finger moves, with local IK and servo doing execution.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
YuanKJing/Jev-as-PolicyPython
YuanKJing/Jev-as-PolicyPython68GitHub stars
Open Thai plus English decision model that answers typed choice, score, and noul questions about a state in one forward pass.
Jev Demos & Experiments#community#jevbest#catalog-import
42GitHub stars
Open harness that renders allowed JSON answers on a diffusion canvas, denoises for a fixed step budget, then picks the top allowed tokens from final logits with no training.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
JoshuaSP/open-jevPython
JoshuaSP/open-jevPython41GitHub stars
Open MIT family of typed-decision models in 4B and 27B with GPU and Apple Silicon builds, answering structured questions in one forward pass with calibrated per-option probabilities.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
caiovicentino/eikosPython
caiovicentino/eikosPython22GitHub stars
Independent open-source local decision model inspired by Jev, with a spark-s1 family plus smaller control models for agent gates, served locally with probabilities and policy actions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
abhishek085/open-spark-jevPython
abhishek085/open-spark-jevPython8GitHub stars
Local server emulating the Jev System One endpoint at POST /v1/systemone with typed noul, choice, and score questions, so the official SDK works with only base_url changed.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
us/jev-localPython
us/jev-localPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.