Jev vs Jev-Omni
Open Gemma-4-12B that scores options on text, images, audio, and video.
Jev 1.13.0
TypeSafe · hosted System One model
77.1Capability · unranked reference
Intelligence 63.6 · calibration 90.6.
Jev-Omni
akhilaaa3 · Open Gemma-4-12B that scores options on text, images, audio, and video
71.3Open-weights Capability #6
Capability #6 at 71.3. The board score is the text measurement.
Key differences
- Where it runs JevHosted APIJev-OmniSelf-hosted merged Gemma-4-12B
- Licence JevProprietary, hostedJev-OmniApache-2.0, following Gemma 4; dataset rights are separate
- Inputs JevText · images (experimental, 1.14)Jev-OmniText, image, audio (30 s), or video (16 frames)
- Capability (v1.6.1) Jev77.1 · referenceJev-Omni71.3 · rank #6
- Intelligence / calibration Jev63.6 / 90.6Jev-Omni55.5 / 87.0
- Name JevTypeSafe’s JevJev-OmniAn independent project. The name is not a TypeSafe release
JevBench v1.6.1, same method
Capability is the mean of intelligence and calibration, ranked only inside Benchmark Heaven’s cost and median-latency caps. Jev 1.13.0 stays on the board as the reference row and is not given a rank. A model can score above that reference and still be left off the numbered list when its estimated cost is too high. Full table ↗
Capability axes, higher is better
Cost $0.029 per 1,000 decisions (estimated). Median latency 0.5 s. Jev’s reference row is $0.032 per 1,000 decisions (not published as a list price), median 0.2 s. Both dollar figures are the benchmark’s, except Jev’s, which Benchmark Heaven marks as not a published list price.
What Jev-Omni is
Jev-Omni is akhilaaa3’s Apache-2.0 classifier on Gemma-4-12B-it. You pass a question and options and get a probability for each option, not a written explanation. The loader accepts an image, an audio file, or a video alongside the text state. It is not published by TypeSafe, and it is not a build of the hosted Jev weights.
On JevBench v1.6.1 it ranks #6 at Capability 71.3. Jev’s text reference is 77.1, with higher intelligence (63.6 vs 55.5) and higher calibration (90.6 vs 87.0). Estimated cost is slightly lower for Jev-Omni ($0.029 vs $0.032 per 1,000 decisions) and the median call is slower (0.5 s vs 0.2 s). The leaderboard measurement is text. The audio and video paths are on the model card.
Read this before you switch
- The shared name causes mix-ups. Jev-Omni weights are not Jev weights.
- Jev leads both axes of the text Capability score.
- Audio is capped at 30 seconds and video at 16 frames on the author’s loader. ffmpeg is required for audio.
When to use which
Choose Jev if
- The input is text and you want the higher Capability reference.
- You want a hosted, versioned API and a name that means TypeSafe’s model.
- Median latency around 0.2 s matters more than a local 12B.
Choose Jev-Omni if
- The decision depends on a photo, a short clip, or a 30-second audio file.
- You want Apache-2.0 Gemma-4-12B weights you can run yourself.
- Estimated cost near $0.029 per 1,000 decisions fits, and you can accept Capability 71.3 on text.
FAQ
Is Jev-Omni a version of Jev?
No. It is an independent Apache-2.0 model by akhilaaa3 on Gemma-4-12B. TypeSafe does not publish those weights. The similar name is the author’s.
Is Jev-Omni better than Jev?
Not on the text board. Capability 71.3 vs 77.1, intelligence 55.5 vs 63.6, calibration 87.0 vs 90.6. It adds audio and video inputs that Jev does not take — Jev 1.14 already reads images experimentally.
Can I send Jev-Omni a video?
The model card says yes: pass media and modality="video". The loader uses 16 frames. Audio is capped at 30 seconds. JevBench did not rank those paths.