JEV COMPARISON · JEVBENCH V1.3.0 · UPDATED SEPTEMBER 22, 2026

Jev vs djev

The fastest decision API, against the best calibrated one.

Jev is built for reliable probabilities and difficult decisions. djev is built for speed and native multimodal input.

On this site

Jev 1.13.0

TypeSafe · System One model

74.4

Leads the benchmark on calibrated probabilities and hard cases.

JevBench v1.3.0 rank #1
vs

Alternative

djev

Maisa · DiffusionGemma decision API

73.0

The fastest model on the board, and the one that can look at a picture.

JevBench v1.3.0 rank #3
Hosted production API
WHERE IT RUNS
Hosted API, free preview
Text only
SUPPORTED INPUTS
Text, images, image options, live camera
0.65s median
RESPONSE TIME
0.24s median
Calibrated probabilities
PROBABILITIES
Documented as experimental; do not threshold directly
74.1% correct
HARDEST TEST CASES
69.5% correct

The benchmark

JevBench v1.3.0, measured the same way

One benchmark ran all 52 systems under one method, so these bars are comparable with each other in a way that vendor-published figures are not.

Composite score

Higher is better

JevBench v1.3.0

#1 Jev
74.4
#2 SemIf
73.1
#3 djev
73.0
#11 OpenJev
66.4
#33 Laya
54.4

The top three finish within 1.4 points of each other, and then the board falls away sharply. The composite hides where systems actually differ, so the two charts below break it apart.

Accuracy by difficulty

Where the decisions separate

Easy

72 straightforward cases

Jev 100%djev 100%

Standard

96 everyday cases

Jev 99%djev 97.9%

Judge

146 evaluation-style calls

Jev 94.5%djev 93.2%

HardHard

220 genuinely ambiguous cases

Jev 74.1%djev 69.5%

Easy and standard decisions separate almost nothing. The hard tier, highlighted, is where these systems stop agreeing.

What the composite leaves out

The composite also weighs a fourth axis: cost, scored from each system's public list price. Jev places last on that axis, while djev is currently free during preview.

See Jev pricing

Every system on the board

The ranking and deployment detail

The capability scores are charted above; this is the ranking and the deployment detail behind them. Jev and djev are highlighted.

ModelRankScoreMeasured latencyRuns on
Jev 1.13.0 (TypeSafe)#174.40.65s median / 0.72s p95Hosted production API
SemIf (Qwen3.5-4B)#273.10.20s raw / 0.55s adjustedSelf-hosted, RunPod GPU
djev (Maisa, DiffusionGemma)#373.00.24s median / 0.31s p95Hosted API, free preview
OpenJev (razorback16)#1166.40.24s raw / 0.63s adjustedSelf-hosted, RunPod GPU
Laya (ModernBERT-large)#3354.40.79s raw / 1.72s adjustedSelf-hosted, CPU

JevBench v1.3.0 scored 52 systems on the same 534 decisions (72 easy, 96 standard, 146 judge, 220 hard) on September 21, 2026, one request at a time.

Feature by feature

Jev and djev side by side

AttributeJev 1.13.0djev
What it isHosted System One model from TypeSafe, version jev-1.13.0Hosted DiffusionGemma decision API from Maisa, in free preview
Question typesNoul, choice, and score — many per call, answered in parallelNoul, choice, and score
InputsText only: a string, JSON object, or arrayText, native images, image options, and live camera frames
Context64k tokens per request, with a 32k state budgetNot published for the hosted preview
CalibrationTrained with RLCD; JevBench calibration score 82.7Maisa documents probabilities as experimental and uncalibrated; JevBench score 65.4
Cost to runMetered in Jev AI credits; see the pricing page$0.026 per 1,000 decisions announced; nothing charged during preview
Open weightsNoThe community djev-dev repo is Apache-2.0; the hosted service is not released
Self-hostingNot availablePossible with roughly one NVIDIA B200, Linux, CUDA 13, and BF16

What each is better at

The practical trade-off

djev is remarkably close in the aggregate. The meaningful difference is calibration, input modality, and how much infrastructure you want to operate.

Where Jev wins

  • Calibration: 82.7 against 65.4.
  • Hard-tier accuracy: 74.1% against 69.5%.
  • A real price, rate limits, and a pinnable version.
  • Focused text decisions with structured output.

Where djev wins

  • Speed: 91.4 against Jev's 83.3.
  • Native image input, image choices, and live camera sampling.
  • Free while the preview lasts.
  • Intelligence close enough that easy decisions rarely separate.

The analysis

The gap is not simply intelligence

In JevBench v1.3.0, djev scores 91.4 on speed against Jev's 83.3. Its intelligence score of 82.7 is close behind Jev's 85.7, so easy and standard decisions rarely show a dramatic difference.

The gap is calibration: 65.4 against Jev's 82.7. Maisa's own documentation describes djev probabilities as experimental and uncalibrated. That matters when code uses confidence to auto-approve, escalate, weight, or route.

The commercial difference is just as clear. djev is a free preview with announced pricing, while Jev is a hosted production API with a version you can pin and a usage model you can plan around.

Operating model

djev trades API spend for GPU ownership when you self-host; Jev is ready as a hosted API.

Input boundary

djev accepts native image and camera input. Jev stays focused on text and structured decisions.

Decision shape

Both are strongest when software needs a typed choice, score, or no/yes judgment.

From the GitHub repository

What djev actually is

djev is a focused, open implementation of typed decisions on DiffusionGemma and vLLM. DiffusionGemma refines answer positions together, then the API reads probabilities for allowed labels and returns a validated response instead of generating and reparsing prose.

The repository adds one-step structured reads, compact canvases, exact label probabilities, reusable prefixes, multimodal attention, and deliberate batching. It does not introduce new model weights.

Read djev-dev on GitHub

Self-hosted

Run the model on your own GPU or rent one when volume justifies it.

Multimodal

Evaluate native image inputs, image choices, and live camera frames.

Direct probabilities

Return typed option probabilities without a full autoregressive output loop.

Reference hardware

The project recommends one NVIDIA B200, Linux, CUDA 13, and BF16.

The djev-dev project is independent and is not affiliated with or endorsed by TypeSafe or Jev AI.

So which should you use?

Choose around your constraints

Pick Jev if

  • Your code branches on the probability, not just the winning label.
  • Inputs include long, messy, or genuinely ambiguous cases.
  • You need forecastable cost and a version held steady across a release.

Pick djev if

  • Your decision is about an image, frame, or scanned document.
  • Latency is the binding constraint and you only need the winning label.
  • You are comfortable building against a preview whose terms may change.

Try it on your data

Try Jev on the cases you are actually arguing about.

A scoreboard is useful for orientation. Your own awkward cases are the better test.

Open the Jev playground

FAQ

Jev vs djev FAQ

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 djev service has not released its hosted implementation.

Is djev free?+

It is free during the preview. Maisa has announced $0.026 per 1,000 decisions for afterwards, but that price is not currently being billed.

Can Jev read images like djev?+

No. Jev accepts text, JSON objects, and text arrays. Convert images to text or structured fields first, or use a model that accepts pixels.

Which is faster, Jev or djev?+

djev, on the hosted APIs measured by JevBench: 0.24s median against Jev's 0.65s. Both are designed for request-time use; neither should be assumed to fit a tight unbatched loop.

Can I use djev probabilities as thresholds?+

Treat them as a ranking signal until you validate them on your own labelled data. Maisa documents the hosted preview probabilities as experimental and uncalibrated.

Figures are quoted from the public sources above as read on September 22, 2026. djev and every other product named on this page belongs to its respective owner.