Two projects · one clear comparison

OmniJev vs Jev-Omni

Similar names. Different backbones, weights and input paths. Find the model that fits the decisions you need to make.

Try native Jev in the playground How to use Jev

These two models are not offered through this site’s API. Try native Jev here for text decisions.

QWEN3.5 FAMILY

OmniJev

Visual decisions across three model sizes.

0.8–4B

Screenshots Video Typed questions

Explore OmniJev tinnel123666888/OmniJev

GEMMA 4 BACKBONE

Jev-Omni

A 12B classifier spanning text and media.

12B

Images Audio Video

Explore Jev-Omni akhilaaa3/Jev-Omni

What actually differs

Both return decision probabilities instead of generated explanations. Their weights and loaders are not interchangeable.

Capability	OmniJev	Jev-Omni
Backbone	Qwen3.5	Gemma 4 12B IT
Released sizes	0.8B, 2B and 4B	12B
Image input	Images, screenshots and image regions	One image per request
Video input	16 timestamped frames in a mosaic	16 sampled frames
Audio path	Sound experiments use spectrogram images	Audio files, capped at 30 seconds
Text input	Instructions and text context with visual input	Text-only or multimodal states
Python loader	mso.infer.MSO1("ckpt", "base")	load_jev_omni()
Setup	Decision checkpoint plus matching Qwen3.5 backbone	Merged weights; loader downloads multimodal components

Read the test, not just the number

Published results, separate evaluations

These panels show selected author-reported results on different datasets. They do not measure which model is more accurate against the other.

OmniJev-4B

LIBERO-10 robot decisions80.7%

1,504 held-out rows

Mind2Web73.3%

1,500 held-out rows

LongVideoBench58.2%

500 validation rows

Author-reported · sources below

Jev-Omni

DecisionBench Medium87.57%

80 scenarios / 293 questions · scenario-weighted

MMAU63.1%

1,000 questions · micro accuracy

MVBench53.1%

14 tasks / 2,786 questions · task-weighted

Author-reported · sources below

No shared head-to-head test is documented in these sources. Dataset composition, averaging methods and runtime setups differ. Test both on your own cases before choosing on accuracy.

Latency needs hardware context

Different GPUs and serving paths: these measurements are not a speed ranking.

OmniJev-4B

294 ms

A800-SXM4-40GB · 12-run medians

One image, one question, 768-token image budget

Jev-Omni

26 ms

Warm H200 · 20-request medians

Image input, optimised backend; preprocessing and network excluded

Which should you evaluate?

Start with OmniJev when…

You want to evaluate 0.8B, 2B and 4B options.
Your workflow asks several questions about a shared screenshot or video state.
You want examples for browser, phone, game or robot decisions.

OmniJev models, benchmarks & setup →

Start with Jev-Omni when…

Your workflow needs its documented audio-file input.
You want to evaluate the Gemma 4 12B classifier.
You can accommodate its documented CUDA setup and roughly 50 GB of FP32 weights before overhead.

Jev-Omni capabilities & setup →

Sources reviewed September 25, 2026: OmniJev repository and Jev-Omni model card. This comparison does not add either model to the Jev AI playground or API.