Model guide · DiffusionGemma decision API · hosted preview · reads images · updated October 7, 2026
djev
Decisions read off a diffusion model in one denoising step.
djev answers typed questions by laying out a small answer canvas with one slot per question, denoising the slots together on Google’s DiffusionGemma, and reading each allowed label’s probability straight off the result. Maisa runs it as a hosted preview at api.djev.dev, and a separate community repository publishes the same approach under Apache-2.0. It reads images and can sample a live camera. Benchmark Heaven dropped it from the v1.6.0 roster, so its newest independent scores are from v1.5.6.
djev is an independent project by Maisa. Jev AI does not host it; this page is a guide to what it is and how to run it yourself.
- Developer
- Maisa
- Base model
- Google DiffusionGemma under vLLM
- Hosted
- api.djev.dev, preview terms
- Open source
- Community djev-dev repo, Apache-2.0
- Inputs
- Text, images, image options, live camera
- Question types
- Noul · choice · score
- Self-hosting
- About one NVIDIA B200, Linux, CUDA 13, BF16
- Last JevBench
- v1.5.6: #22 of 110 at 64.2
djev on JevBench v1.6.0
One benchmark measured every system under one method and ranked 92. The full board is on the JevBench results page.
djev is not on the JevBench v1.6.0 roster, so there is nothing current to chart against Jev. Its last published scores, from v1.5.6, are quoted in the text below and are not comparable with v1.6.0 figures.
What djev is
A normal language model writes an answer one token at a time; a diffusion model starts from noise across the whole output and refines it. djev exploits that by laying out a small answer canvas with one slot per question, denoising the slots together, and reading the probability of each allowed label off the result. There is no prose to parse and no JSON to repair.
Maisa runs the hosted service at api.djev.dev as a preview whose terms can change; open-sourcing the service is planned but has not happened. The community djev-dev repository publishes the same approach under Apache-2.0 for self-hosting and recommends a single NVIDIA B200 on Linux with CUDA 13 and unquantised BF16 weights. It reports sub-100ms warm text responses from internal measurement, while noting those figures are not benchmarks for the current BF16 release and do not apply to image requests.
The distinctive capability is multimodal: native image inputs, images as the options of a choice question, and live camera sampling. Jev is text only.
Its last JevBench scores
JevBench v1.6.0 re-measured every system with a reproducible recipe on a new pool and djev is not on the resulting roster, so there is no current independent score. In v1.5.6, the last release that measured it, djev was level with Jev on intelligence, 72.3 against 72.0, and ahead on yes/no questions, 68.2 against 48.2 above chance, but ranked #22 of 110 at 64.2 against Jev’s #3, because the cost estimate for running DiffusionGemma was its weakest axis and the composite punishes one weak axis hard. Calibration was 80.4 against Jev’s 88.0, and Maisa’s own documentation describes djev’s probabilities as experimental and uncalibrated.
Maisa also entered djev (thinking). In v1.5.6 its calibration was 95.7 and its intelligence 77.3, both above Jev in that release, but thinking costs far more compute per decision and it placed #62 at 17.4. Neither configuration appears in v1.6.0.
Use djev
Call Maisa’s hosted preview with an API key from djev.dev, or self-host the community implementation on a B200-class machine.
# Hosted preview (terms can change): see https://api.djev.dev for the current endpoint and auth
# Self-hosted, community implementation:
git clone https://github.com/Davipar/djev-dev
cd djev-dev
# Follow README.md: Linux, CUDA 13, unquantised BF16 DiffusionGemma weights under vLLM.
# Then send the state, typed questions and, optionally, images.- Treat the probabilities as a ranking signal rather than thresholds until you have validated them on your own labelled data; Maisa documents them as experimental.
- The recommended self-hosting machine is a single NVIDIA B200, which is a serious GPU.
- Image requests are slower than the sub-100ms warm text figure the repository reports.
Limits to plan around
From the project’s own documentation and JevBench v1.6.0.
- Preview terms that can change, and no current independent benchmark score.
- Probabilities documented as experimental and uncalibrated.
- Context limit not published for the hosted preview.
- Self-hosting needs a B200-class GPU and BF16 weights.
Measure it against Jev on your own cases
If you want to see the calibration difference rather than read about it, run both on your own awkward cases. The Jev AI playground gives you Jev 1.13.0 with no setup; every answer comes back with its full probability distribution and a confidence value, which is the thing you are actually comparing.
djev FAQ
What is djev?
Maisa’s decision model built on Google’s DiffusionGemma. It answers noul, choice and score questions by denoising an answer canvas in one step and reading label probabilities off the result, and it accepts images and live camera frames. It runs as a hosted preview at api.djev.dev.
Is djev open source?
Partly. The community djev-dev repository is Apache-2.0 and shows how to run DiffusionGemma decisions on vLLM. Maisa’s hosted service says open-sourcing is planned but has not released its implementation.
Why is djev not on JevBench v1.6.0?
Benchmark Heaven’s v1.6.0 release re-measured every system with a reproducible recipe on a new decision pool, and djev is not on the resulting roster of 127 systems. Its last published scores are from v1.5.6.
Can I use djev probabilities as thresholds?
Maisa’s documentation calls them experimental and uncalibrated. In JevBench v1.5.6 its calibration scored 80.4 against Jev’s 88.0 in that release. Validate on your own labelled data first.
Is djev the same as Jev-Omni or OmniJev?
No. All three read images, but djev is Maisa’s hosted DiffusionGemma model, Jev-Omni is a community Gemma 4 12B fine-tune and OmniJev is a Qwen3.5-based project from a Beijing research group. This site has separate guides for the other two.
More model guides
Every figure on this page is quoted from a published source and was read on October 7, 2026.
- OpenJevOne name, ten projects. Find yours, then run the drop-in.
- Winnow-12BOne local model that decides, chats and looks at a screenshot.
- Imajev-4BJev’s request format, open weights, and two photos per request.
- Jev 1.13.0The hosted model every project here is measured against.
- Jev-OmniGemma 4 12B decision model with image input.
Sources
djev and every other product named on this page belongs to its respective owner and is not affiliated with Jev AI. Figures are quoted from the sources above as they read on October 7, 2026 and can change.