imajev
Small open vision-language decision models (2B, 4B, 9B) that answer typed questions about photos and records with a trained can't-tell.
Overview
imajev is a family of open Apache-2.0 decision models at 2B, 4B and 9B parameters, built by Mohit Garg. They read the photos, records and text a business already has and answer in the options the caller sets, with a probability on each and an explicit unknown answer so the application can act when sure and hand the rest to a person. Target cases include refunds, returns, ticket routing, listing-versus-photo checks and parts inspection.
The server speaks TypeSafe's POST /v1/systemone contract plus an images field, unknown_probability and abstained. The README cites external leaderboards captured on 28 September 2026: imajev-4b first of 91 on JevBench v1.4.2.2 (67.37 versus 63.29 for Jev 1.13.0), first of 49 on Image JevBench v0.1.3, and third of 56 on DecisionBench. A showcase request ran in 1.15 s on a Mac Studio with four option orders averaged and calibration applied. A live Hugging Face Space demo and a technical report are linked.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which field of `listing` does this photo contradict? | listing.color · listing.product_type · none of these |
| Other | Can the model tell at all? (trained unknown probability, abstained flag) | — |
How it works
- Send photos, a record and typed questions to the local server
- Model scores each allowed option, averaging over option orders
- Calibration file is applied and an unknown probability returned
- App acts when confident or routes to a person when unknown is high
Strengths
- Runs locally (MLX or one GPU), so photos and customer data stay in-house
- Trained can't-tell answer supports safe abstention
- README reports top leaderboard positions against Jev 1.13.0
- Drop-in Jev request shape with images added
Caveats
- Leaderboard ranks are dated snapshots that move as systems are added
- State limit (32 KB) is smaller than hosted Jev's
- Author also sells consulting, so results are self-reported via screenshots
Getting started
Try the Hugging Face Space at huggingface.co/spaces/mohit67890/imajev, or run the local server and POST to http://127.0.0.1:8765/v1/systemone.At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- local server (MLX or PyTorch)
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Open Jev-style typed-decision model that also takes images: photo + app state +…
wh000wh000/awesome-jev-liveSmall open models at 2B, 4B and 9B that read the photos, records and text a business already has and answer in the options you set, with a probability on each and an explicit "can't tell" that sends the rest to a person.
MrJev/awesome-jev
Frequently asked questions
What does imajev decide?
Typed questions about photos and records, such as which listing field a photo contradicts, with an explicit unknown option.
Does it need a TypeSafe key?
No. It is an open model you run locally with the same API shape as Jev.
Which size should I use?
The project recommends imajev-4b as the default.
Related projects
Perplexity's Apache-2.0 multimodal decision model on Qwen3.8-27B, reporting 61.56 on the Decision Index versus 57.9 for Jev.
ServiceproductionPythonHead-to-head eval page where TrustedRouter's gev-1.0 outperforms Jev on the same JevBench questions and also accepts images.
Articlediscussion- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPython