Imajev-4B
28.7rank #26 · Capability #15
Model guide · Open decision model · Qwen3.5-4B · reads photos · updated October 7, 2026
Imajev-4B
Jev’s request format, open weights, and two photos per request.
Imajev is a family of open decision models that answer the same yes/no, choice and score questions as Jev, from a LoRA adapter on Qwen3.5. The 4B is the recommended size. It keeps Qwen’s vision tower, so a request can carry up to two images alongside the text, and every answer comes with a trained probability that the evidence cannot settle the question. This guide covers what it is, what it needs and where it stands on JevBench; the comparison page argues the case against Jev.
Imajev-4B is an independent project by Mohit Garg. Jev AI does not host it; this page is a guide to what it is and how to run it yourself.
- Developer
- Mohit Garg (independent)
- Sizes
- 2B · 4B · 9B
- Base model
- Qwen3.5, vision tower frozen
- Licence
- Apache-2.0 adapters, code and base
- Inputs
- Text plus up to two images per request
- State limit
- 32 KB, about 8k tokens
- Questions
- 1 to 8 per request, 2 to 254 options
- Runs on
- Apple silicon (MLX) or one CUDA GPU
Imajev-4B on JevBench v1.6.0
One benchmark measured every system under one method and ranked 92. The full board is on the JevBench results page.
Jev 1.13.0
71.1rank #2 · Capability #3
IntelligenceHow often it picks the right answer, half from sealed decisions
CalibrationWhether 0.8 really means about 80%
SpeedMeasured response time
Open half 35.0 · sealed half 34.2; chance-corrected by request type: choice 64.4, yes/no 0.3, score 39.1. The composite also weighs a fourth axis, cost, which these bars leave out. Self-hosted latency is adjusted by the benchmark to approximate production load.
What Imajev-4B is
Each Imajev size is a LoRA adapter plus a 255-code decision readout on a Qwen3.5 model. It reads the state, the images and each question, then takes the probability of every option from one position in a single forward pass. No text is generated, so there is no answer to parse.
Its server accepts the same /v1/systemone request and response as TypeSafe’s Jev, and adds three things: image inputs, an unknown probability on every answer and an abstained flag. The README says text-only requests written for Jev work unchanged.
Training used about a million decisions in four stages, labelled by people or by open-weight teacher models. The author states that no Jev outputs, no paid-API outputs and no JevBench items were used, and that JevBench items were screened out with an 8-gram check.
Where it stands on JevBench
JevBench v1.6.0 measured the released 4B adapter on an RTX 5090, with one option order and the shipped calibration file. It ranks #26 of 92 at 28.7, with intelligence 34.6 and calibration 89.2; Jev is #2 at 71.1. The weak spot is yes/no questions, where it scores 0.3 above chance on the new pool; on choice questions it is closer, 64.4 against Jev’s 76.0.
It led the earlier v1.4.2.2 release, and the drop is the benchmark changing rather than the model: v1.5 made the sealed half count for half of intelligence and v1.6.0 re-measured everything on a new pool. JevBench is text-only. On Benchmark Heaven’s separate Image JevBench v0.1.5, which does test photos, Imajev-4B is second at 76.4, ahead of Jev-Omni.
Run Imajev-4B
The repository ships a server and the adapters; the 4B needs a 9.3 GB base-model download. Follow its README for the MLX or CUDA install, then send Jev-style requests to it.
git clone https://github.com/mohit67890/imajev
cd imajev
# Follow README.md: install for MLX (Apple silicon) or CUDA,
# then download the 4B adapter from huggingface.co/mohit67890/imajev-4b
# and start the server. Text-only Jev requests work unchanged:
curl http://localhost:8000/v1/systemone \
-H "Content-Type: application/json" \
-d '{"state": "Order #4411: customer says the box arrived crushed.",
"questions": {"damaged": {"type": "noul", "instructions": "Does the customer report physical damage?"}}}'- The author measures about 0.1s per question raw on one H100 and about 0.35s as shipped, averaging four option orders.
- Keep the shipped calibration file: without it the model is over-confident on hard items, by the author’s own account.
- An empty field can be read as “no” rather than unknown, so send unknowns explicitly.
Limits to plan around
From the project’s own documentation and JevBench v1.6.0.
- English only.
- State capped at 32 KB, roughly 8k tokens, against Jev’s 64k.
- Barely above chance on yes/no questions in JevBench v1.6.0.
- ImajevBench, the author’s own image benchmark, uses AI-generated images that have not had a human audit.
Measure it against Jev on your own cases
Imajev-4B trails Jev on text but does something Jev cannot with photos, so the useful test is your own cases. Run the text ones on Jev here with no setup and keep the answers. If your decisions need a photo, this site also hosts Jev-Omni, the other open decision model that reads images.
Imajev-4B FAQ
What is Imajev-4B?
An open, Apache-2.0 decision model by Mohit Garg: a LoRA adapter and decision readout on Qwen3.5-4B that answers Jev-style noul, choice and score questions about text and up to two images, returning probabilities instead of generated text.
Is Imajev-4B made by TypeSafe?
No. It is an independent project that mirrors Jev’s request contract. Its author states that no Jev outputs were used in training.
Can Imajev read images?
Yes. A request can include up to two images, for example a reference photo and a target, alongside the state and questions. Jev itself is text only.
How does Imajev-4B rank on JevBench?
#26 of 92 at 28.7 in JevBench v1.6.0, against Jev at #2 with 71.1. On the text-only pool it is barely above chance on yes/no questions. On Image JevBench v0.1.5 it is second at 76.4.
What hardware does Imajev-4B need?
Apple silicon through MLX or one CUDA GPU, plus a 9.3 GB base-model download for the 4B. JevBench measured it on an RTX 5090.
Can I use Imajev-4B on Jev AI?
No. This is a model guide; Jev AI hosts Jev, Laya, Mercury Decide and Clef. If your decisions need a photo, Jev-Omni is the other open image-capable decision model this site documents.
More model guides
Every figure on this page is quoted from a published source and was read on October 7, 2026.
- JevK5The fastest open 4B on the board, with the weights published.
- decider-4bMillisecond decisions from 8.4 GB of open weights, on TypeSafe’s wire format.
- OpenJevOne name, ten projects. Find yours, then run the drop-in.
- Jev 1.13.0The hosted model every project here is measured against.
- Jev-OmniGemma 4 12B decision model with image input.
Sources
Imajev-4B and every other product named on this page belongs to its respective owner and is not affiliated with Jev AI. Figures are quoted from the sources above as they read on October 7, 2026 and can change.