Jev vs djev

The analysis

Where Jev and djev really differ, and why the bars above land where they do.

Benchmark Heaven’s v1.6 release re-measured every system with a reproducible recipe on a new pool and dropped djev from the roster altogether, so there is no current independent score for it. Every djev figure on this page is from v1.5.6, the last release that measured it, and is quoted beside Jev’s v1.5.6 figures so that both sides come from the same run. Jev’s own current score is 71.4 on board v1.7.48, where it is the open-weights board’s unranked reference row and #15 of 48 on the API board; those numbers are not comparable with the v1.5.6 ones below.

In that run, djev was level with Jev on intelligence, 72.3 against 72.0, and well ahead on yes/no questions, 68.2 above chance against 48.2, but on that 110-system board it ranked only #22 at 64.2 against Jev’s #3 at 72.1, because the cost estimate for running DiffusionGemma was the weakest axis on its row, 48.2 against Jev’s 54.7, and the composite punishes one weak axis hard. The gap that matters day to day is calibration: 80.4 against Jev’s 88.0, and djev’s own documentation describes its probabilities as experimental and uncalibrated. That matters as soon as your code does anything other than take the top label — auto-approving above 0.9, escalating below 0.6, weighting a score. Jev also led on choice questions, 86.7 against 74.7, and score questions, 81.1 against 74.1.

The other gap is maturity, and the missing v1.6 row is part of it. djev is still a preview whose terms can change, open-sourcing is planned but has not happened, and it is now absent from the leaderboard everything else here is measured on. Jev is generally available, with published rate limits and a version you can pin.

What djev actually is

djev, the DiffusionGemma decision API from Maisa, by Maisa.

djev is built on Google’s DiffusionGemma running under vLLM. 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 straight 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. A separate community repository, djev-dev, publishes the same approach under Apache-2.0 for self-hosting; it recommends a single NVIDIA B200 on Linux with CUDA 13 and unquantised BF16 weights, which is a serious machine. The repository 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. djev accepts native image inputs, lets you use images as the options in a choice question, and can sample a live camera. Jev is text only: a string, a JSON object or an array. If your decision is about a photograph, Jev cannot make it and djev can.

One djev is better calibrated than Jev

Maisa entered a second configuration, djev (thinking), and it complicates the section above in Maisa’s favour. In JevBench v1.5.6 its calibration score was 95.7, well clear of Jev’s 88.0 in that release, and its intelligence 77.3 against 72.0.

It still placed only #62 on that board, at 17.4. Thinking takes far more compute per decision, its cost score was 28.1, and the harmonic mean turns one weak axis into a low composite. If the section above left you thinking DiffusionGemma cannot produce trustworthy probabilities, the thinking variant is the counterexample — and it is Maisa’s own. Neither configuration appears on board v1.7.48.

Jev vs djev FAQ

Other Jev comparisons

Every number on these pages is quoted from a published source and was read on October 31, 2026.

• Jev vs Imajev-4BA 4B model that reads photos, against the model it copies on text.
• Jev vs decider-4bLevel on speed, far behind on the answers.
• Jev vs JevK5A fast open Qwen3.5-4B system, and far behind Jev on the answers.
• Jev vs CygnetNo training at all, forty-sixth on the open-weights board, and behind Jev on every accuracy axis.
• Jev vs decisioTenth on Capability on its biggest base, and it trains nothing of its own.
• Jev vs metask-jev-rainMore accurate answers than Jev on the benchmark, and two of three rows above it on the composite.
• Jev vs WinnowOne local model for decisions, chat and images, against a hosted decision model.
• Jev vs SemIfA fast open rebuild on your own GPU, and where it still falls short.
• Jev vs OpenJevOne name, ten projects. Here is every one, and how the drop-in compares.
• Jev vs LayaA model you fine-tune on your data, against one that works out of the box.
• Jev vs Mercury DecideTwo hosted decision models that take the same request, one place apart.
• Run Jev locally?The open alternatives in one table, with licences and hardware.
• Jev vs LLMHow a decision model differs from a chat LLM and a BERT classifier.

Sources

• Benchmark Heaven — JevBench board v1.7.48 (open weights)
• Benchmark Heaven — JevBench API board
• djev-dev — DiffusionGemma structured decisions (Apache-2.0)
• TypeSafe documentation (Jev model card)