Imajev-4B
Photos checked against your record, with a trained “can’t tell”.
4B
Two photos Can’t tell Typed questions
Imajev-4B against Jev mohit67890/imajev · also 2B and 9BTwo open image decision models · updated October 9, 2026
Imajev-4B vs Jev-Omni
Both read a photo, take your typed questions and return a probability for every option instead of prose. Each is ahead on a different leaderboard: Imajev-4B leads Image JevBench, and Jev-Omni ranks far higher on the text JevBench. Here is why, and which one fits your decisions.
Jev-Omni is hosted in our Playground in beta. Imajev-4B is not offered on this site, and neither model is available through our public API.
Jev-Omni
One classifier for text, images, audio and video.
12B
Images Audio Video
Explore Jev-Omni akhilaaa3/Jev-Omni · hosted hereMeasured by the same third party
Two leaderboards, opposite orders
Benchmark Heaven runs both boards. On images, Imajev-4B is the top open-weights system and well clear of Jev-Omni on both accuracy axes, while Jev-Omni still answers faster. On text, Jev-Omni ranks thirty-seven places higher, #16 against #53, and scores 21 points higher on intelligence.
| Measure | Imajev-4B | Jev-Omni |
|---|---|---|
| Image JevBench v0.3.0 · images, 1,500 decisions | ||
| Rank and composite | #2 of 52 · 63.7 | #7 of 52 · 32.8 |
| Capability rank | #1 · 68.0 | #8 · 54.3 |
| Intelligence | 49.9 | 37.5 |
| Calibration | 86.1 | 71.0 |
| Adjusted median response | 0.37s | 0.32s |
| JevBench v1.6.1 · text, 1,500 decisions | ||
| Rank and composite | #53 of 135 · 28.7 | #16 of 135 · 67.7 |
| Capability rank | #35 · 61.9 | #16 · 71.3 |
| Intelligence | 34.6 | 55.5 |
| Calibration | 89.2 | 87.0 |
| Sealed half | 34.2 | 51.9 |
Version 0.3.0 is a fresh start, not an update: a new pool of 2,441 items on 2,160 images, drawn as 300 public and 1,200 sealed decisions, of which 713 items are real and 1,728 synthetic. 54 systems were measured and 52 ranked; two more keep a separately dated v0.1.5 score and are not ranked on the v0.3 scale. No v0.3.0 score can be compared with a v0.1.5 one, and the per-system public and sealed item counts the old release published are not part of this one. Wity-1, the closed API that led v0.1.5, was re-measured on the new pool on 7 October and holds first place again at 71.2, which puts Imajev-4B second and makes it the highest-placed system whose weights you can download.
Read the test, not just the number
The Imajev author’s own image test
ImajevBench is the Imajev author’s benchmark: 279 test questions about photos, records and text, 21 of which have no answer the evidence supports. It is a preview with AI-generated images and no human audit yet.
83.9% vs 78.5%
Imajev-4B against Jev-Omni, run by the Imajev author with Jev-Omni’s own predict(). The author’s significance test, run on an earlier Imajev release, found a gap of this size not significant.
18 vs 0 of 21
The gap comes from the “can’t tell” questions, which Jev-Omni has no output for. On answerable questions the author found Jev-Omni slightly ahead of an earlier Imajev release, and better calibrated.
What actually differs
Both return decision probabilities instead of generated explanations. Their inputs, limits and serving paths are not the same.
| Capability | Imajev-4B | Jev-Omni |
|---|---|---|
| Backbone | Qwen3.5-4B with a LoRA adapter and decision readout; also 2B and 9B | Gemma 4 12B IT, fine-tuned and merged |
| Images | Up to two per request, such as a reference and a target | One per request |
| Audio and video | Not supported | Audio up to 30 seconds; video as 16 sampled frames |
| Can’t tell | A trained unknown probability and an abstained flag on every answer | No abstain output |
| Request format | TypeSafe-style /v1/systemone, plus images | Its own Python loader and predict(); hosted here in the Playground |
| Options | 2 to 254 per question, 1 to 8 questions per request | Head accepts 256; best supported at 20 or fewer |
| Hardware | Apple silicon through MLX or one CUDA GPU; 9.3 GB base weights | A CUDA GPU; about 24 GB of BF16 weights |
| Licence | Apache-2.0 adapters and code on an Apache-2.0 base | Apache-2.0 weights following Gemma 4; dataset rights separate |
| Try it | The author’s Hugging Face demo, or run it yourself | Hosted in this site’s Playground (sign in), or run it yourself |
Which should you evaluate?
Start with Imajev-4B when…
- You compare a photo with a record: a listing, a return, a part against a known-good one.
- You want the model to say it cannot tell, and send those cases to a person.
- Most of your decisions are about images, where it leads every open-weights system on the image board.
- You need to run it locally, including on a Mac.
Start with Jev-Omni when…
- Your decisions involve audio or video, not only photos.
- Many of your decisions are text-only, where it scores far higher.
- Response time on image decisions matters most.
- You want to try it right now without a GPU, in this site’s Playground.
Imajev-4B vs Jev-Omni FAQ
Which is better on images, Imajev-4B or Jev-Omni?
Imajev-4B, and by a wider margin than before. On Image JevBench v0.3.0 it is second of 52 at 63.7, behind only the closed API Wity-1, against Jev-Omni’s 32.8 at #7, with intelligence 49.9 against 37.5 and calibration 86.1 against 71.0. It also leads that board’s Capability ranking at 68.0, with Jev-Omni eighth at 54.3. Jev-Omni answers faster. On the Imajev author’s own ImajevBench, Imajev-4B scores 83.9% against 78.5%.
Which is better on text?
Jev-Omni, by a wide margin. On the text JevBench v1.6.1 it is #16 of 135 at 67.7, with intelligence 55.5, and sixteenth on the headline Capability ranking; Imajev-4B is #53 at 28.7, with intelligence 34.6 and thirty-fifth on Capability.
Can either model say it cannot tell?
Imajev-4B can: every answer carries a trained unknown probability and an abstained flag. Jev-Omni has no abstain output, so on questions whose honest answer is “can’t tell” it always picks one of the options.
Can I try them without a GPU?
Jev-Omni is hosted in this site’s Playground: sign in and run text, image, audio or video decisions. Imajev-4B has a demo on Hugging Face Spaces run by its author, and it also runs locally on Apple silicon.
Are these made by TypeSafe?
No. Imajev is an independent project by Mohit Garg and Jev-Omni an independent model by akhilaaa3. Both borrow Jev’s typed-decision idea; TypeSafe’s Jev itself reads text only.
Sources read October 9, 2026: JevBench v1.6.1, Image JevBench v0.3.0, Imajev repository and Jev-Omni model card. Imajev and Jev-Omni belong to their respective authors and are not affiliated with Jev AI or TypeSafe.