mohit67890/imajev

Open Jev-style typed-decision model that also takes images: photo + app state + typed questions in, calibrated probabilities out, locally.

📸 imajev – Decision‑making models that read photos and business records

What it is – A family of small, open‑source multimodal models (2 B, 4 B, 9 B parameters) that take a photo (or two) plus structured text data (e.g., a product listing, a support ticket, an invoice) and return typed answers (choice, yes/no, score) together with a calibrated probability and an explicit can’t‑tell flag. The models are built on the Qwen‑3.5 series, frozen vision towers, and a LoRA adapter that learns to output decisions in the same JSON contract used by TypeSafe’s Jev system.

🎯 Core idea

Business‑grade decisions – automate high‑volume, well‑defined actions (e.g., “does the photo match the listing?”) while deferring uncertain cases to a human.
Probabilistic, abstaining output – each answer includes a probability for every option you supplied and an unknown_probability. If the model is not confident enough, you can set a threshold and let a person handle the request.
Runs locally – weights are Apache‑2.0 licensed and can be run on a single GPU or an Apple‑silicon Mac (MLX). No data leaves your network.

🚀 Highlights (from the README)

Feature	Why it matters
Photo + record – up to two images per request (reference + target) and up to ~32 KB of structured data.	Lets you verify a shipment against a catalog, check a part against a known‑good reference, etc.
Typed JSON contract – same request/response format as Jev (choice, `noul` score, `unknown_probability`).	Easy drop‑in replacement for existing text‑only decision pipelines.
Trained can’t tell – the model learns when it truly doesn’t know (18/21 unknown items on ImajevBench).	Reduces hallucinations and avoids costly wrong actions.
Small, local – 2 B–9 B parameters, ~0.1 s raw inference on an H100, ~0.35 s with shipped calibration.	Fits on a single GPU or a modern Mac; low latency for real‑time workflows.
State‑of‑the‑art benchmark scores – #1 on JevBench, Image JevBench, and #3 on DecisionBench, beating many multi‑hundred‑billion‑parameter systems.	Demonstrates competitive quality despite the modest size.

📊 Performance snapshot (official leaderboards, Sep 2026)

JevBench v1.4.2.2 – 1️⃣ / 91 models, score 67.4 (4 B).
Image JevBench v0.1.3 – 1️⃣ / 49 models, score 76.4 (4 B).
DecisionBench (English) – 3️⃣ / 56 models, score 79.65, ahead of GPT‑5.6 Luna.
ImajevBench (image‑text decision benchmark) – 83.9 % accuracy (4 B), 18/21 unknown correctly flagged.

🛠️ How to try it

Clone & install

git clone https://github.com/mohit67890/imajev && cd imajev
python3.11 -m venv .venv && source .venv/bin/activate
pip install -e "[serve,mlx]"   # use "torch" extra on non‑Apple hardware

Download a model (e.g., the 4 B version)

python scripts/download_model.py --model 4b   # pulls Qwen‑3.5‑4B and the LoRA adapter
hf download mohit67890/imajev-4b --local-dir adapters/imajev-4b

Run the local server

PYTHONPATH=src:scripts python scripts/playground/server.py \
  --model-bundle artifacts/model-qwen4b.json \
  --adapter adapters/imajev-4b/mlx \
  --calibration adapters/imajev-4b/calibration.json \
  --model-name imajev-4b --port 8765

Call the API (example with a product listing and a photo)

curl -s http://127.0.0.1:8765/v1/systemone \
  -F 'request={"state":{"listing":{"title":"Blue mug","color":"blue"}},
            "questions":{"matches":{"type":"noul","instructions":"Does the photo show the listed item?"}}}' \
  -F image=@photo.jpg

The response contains the chosen option, per‑option probabilities, confidence, and unknown_probability.

A live demo is hosted on Hugging Face Spaces: https://huggingface.co/spaces/mohit67890/imajev.

📂 Repository layout (high‑level)

src/ – core library that parses the request, runs the model, and formats the JSON response.
scripts/ – utilities for downloading weights, launching the demo server, and the playground apps that showcase real‑world scenarios (business checks, text‑only, wardrobe, stylist, tracing pad).
docs/ – assets, benchmark screenshots, and the technical report.
artifacts/ – model‑bundle JSON that tells the server which base model and LoRA to load.

🏷️ License

Apache‑2.0 – both the code and the model weights are open‑source, allowing commercial use and modification.

🤝 Who built it?

Created by Mohit Garg, an AI consultant focused on decision‑automation for operations, support, and back‑office teams. More info: https://mohitgarg.ai.

📌 When to use imajev?

Retail / marketplaces – verify that a product photo matches a catalog entry, automate return validation.
Manufacturing / field service – compare a shipped part to a reference image to detect damage.
Customer support – route tickets, flag urgency, or validate refund eligibility based on attached screenshots.
Trust & safety – quickly reject content that violates policy, ask the user for a clearer image when uncertain.
Any high‑volume, low‑risk decision where a clear set of answer options can be defined and you want the model to abstain when unsure.

📈 Bottom line

imajev delivers a practical, locally‑runnable multimodal decision engine that bridges the gap between pure‑text LLMs (like Jev) and heavyweight vision APIs. Its small footprint, calibrated can’t‑tell behavior, and strong benchmark rankings make it a solid choice for enterprises that need fast, trustworthy automation on their own hardware.