Jev alternatives
djev
The diffusion route — and the thinnest evidence in this catalogue
Quick answer
djev is worth following, not adopting on faith. It is the diffusion-faction entry — a DiffusionGemma 26B-A4B model read as a canvas rather than decoded token by token — and it is one of two routes cited for accepting image state. Its published JevBench score fell from 74.3 on one board to 52.23 on a later one with no new measurements, and there is no clean head-to-head against Jev at all.
djev belongs to the smallest and most interesting faction: instead of reading logits or training an encoder, it treats a diffusion language model as a canvas, leaves the answer slots blank and fills them all at once. It ships no weights of its own and its benchmark scores move a great deal between board versions, so this is a page about a promising direction with thin evidence rather than a recommendation.
Fill answers with diffusion
Leave the answer slots blank and fill them all at once, the way a diffusion model completes an image, instead of decoding token by token.
Search aliases
Key specs
- Licence
- Apache-2.0 per the repository
- Author
- Maisa
- Backbone
- DiffusionGemma 26B-A4B, read as a canvas
- Size
- No weights shipped by the project itself
- Latency
- 76.87 ms p50 on the older quantized configuration
- Wire format
- See repository
- Install
- See repository
djev and the JevBench board problem
This is the single most important caveat on this page. JevBench re-scored the same systems twice with no new measurements, and the movement was severe — which is why the same project can be described as third place or eighth place depending on which post you read.
| djev and the JevBench board problem | djev | Jev |
|---|---|---|
| JevBench, oldest board | 74.3 — third place | 75.3 — first place |
| JevBench v1.3.0 | 73.0 — third place | 74.4 — first place |
| JevBench v1.4.2 | 52.23 — eighth place | 63.29 — second place |
| Clean head-to-head with Jev | None published | — |
| Approach | Diffusion: answer slots filled in one pass | Single forward pass, non-generative |
When to consider djev
Consider it when the decision depends on an image, or when you are researching diffusion-based decision models rather than shipping one. The faction is genuinely different from the logit-reading crowd: OpenJev and LocalJev both build on this route, and LocalJev’s README is admirably clear that its own bridged version is wire-compatible but not mathematically equivalent. If you need image-aware typed decisions today, expect to do your own evaluation.
When to walk away
Do not pick djev because a blog post called it third on JevBench. That was the oldest of three boards; the current board puts it at 52.23, and the sealed-accuracy column — the part that measures performance on items that were not public — is where this whole category collapses. There is also no published comparison between djev and Jev on controlled inputs, so any claim about how it stacks up is inference rather than measurement.
Evaluating djev responsibly
Diffusion route checklist
-----------------------
[ ] Model: DiffusionGemma 26B-A4B (no djev weights are shipped)
[ ] Runner exposes the structured-read primitives:
- seeded diffusion canvases
- read-only denoising
- selected-token logits / logprobs
Without these, probabilities are "self-reported" by the model
through an OpenAI-compatible endpoint — wire-compatible,
not mathematically equivalent.
[ ] Record which JevBench version any score came from.
74.3 (oldest board) -> 73.0 (v1.3.0) -> 52.23 (v1.4.2)
[ ] Compare on your own labelled decisions, including images
if that is why you are here.djev questions people actually ask
Why did djev’s score drop from 74.3 to 52.23?
Not because the model changed. JevBench re-scored the same systems under new rules — chance-corrected intelligence and a near-chance penalty on one version, a sealed decision set and harmonic mean on a later one — with no new measurements. Movement was brutal across the board: one project went from 67.0 to 40.6 and another from 62.4 to 27.4. Always state which board a number came from.
Does djev ship weights?
No. The project provides the method, not a checkpoint — it reads a DiffusionGemma 26B-A4B model as a canvas. That means your results depend heavily on the runner: you need seeded diffusion canvases, read-only denoising and selected-token logits, or you end up with a chat endpoint reporting its own probabilities.
Can djev handle images?
It is one of two routes described as accepting image state, and sources disagree about which is primary — one widely-read comparison credits razorback16’s OpenJev as the only project answering questions about images, while other write-ups point at djev. Treat the image story as promising and unverified, and test it yourself if that is the reason you are looking.
What is the sealed-accuracy column and why does it matter?
It measures performance on decisions that were not public when the systems were scored, which is the only defence against benchmark contamination. On the current board the best public accuracy runs in the eighties while sealed accuracy lands around the thirties — including for Jev itself. That gap is the most honest number in this whole ecosystem, and it is why we call every comparison table here a hypothesis.
Sources
- Davipar/djev — repository
- DiffusionGemma — the model this route reads
- JevBench — board versions and rescoring
- LocalJev — why a diffusion bridge is not logit reading
Every comparison number on this page is a third-party published figure or a read of a public repository — not a benchmark we ran. We have not tested TypeSafe Jev itself, and its customer agreement forbids using its outputs to build similar products.