AI product Open source
imajev is an open, locally runnable multimodal typed-decision model and repository maintained by Mohit Garg. It accepts up to two images, application state, and typed questions with user-defined answer options, then returns probabilities for each option together with an unknown probability, confidence, and abstention status. Example tasks include checking a product listing against its photo, comparing a returned item with a reference image, routing support tickets, and evaluating records against policies.
Rather than generating an answer, imajev binds each option to a decision code and reads the model's logits at a decision position through a float32 output head. Its vision tower is frozen while LoRA adapters are trained on the language-model projections; the released 2B, 4B, and 9B variants are based on Qwen3.5 models. Unknown is trained as a first-class outcome for missing, contradictory, or out-of-scope evidence, allowing applications to route low-confidence cases to a person. The repository provides a local server and API with MLX and PyTorch paths, calibration files, playground applications, training scripts, and an ImajevBench evaluation suite.
The code, adapters, and stated base models are distributed under Apache-2.0. The README describes English-only operation and limits of up to two images, 32 KB of state, eight questions per request, and 254 options; it also cautions that calibration and testing on application-specific data are needed before setting automation thresholds.
1 use taken from transcripts — each links to the moment in the video.
A vision system that compares images or checks a product listing against an image, returning probabilities for potentially incorrect fields and handling visual questions.
1 in the library.