Jev directory
CommunityJev Models & Reimplementations
Imajev 4B
Mohit Garg's Apache-2.0 vision-language decision model on Qwen/Qwen3.5-4B, the recommended default of the imajev family: it reads photos against records and text, returning calibrated probabilities and a trained can't tell.
- Reading time
- 3 min
- Published by
- Community
- Author
- Mohit Garg
- Last verified
- Oct 2026
Highlights
- Recommended default of the family: 83.9% on ImajevBench against the 9B's 82.1% at under half the size, and the only size trained on phase-3 hard data.
- #1 of 91 on JevBench v1.4.2.2 (67.4 composite) and #1 of 49 on Image JevBench v0.1.3 (76.39); #3 of 56 on DecisionBench at 79.65.
- Trained on 72k photo-vs-record and two-photo decisions; every answer carries a probability for unknown plus an abstained flag.
- At a 90% threshold it automates 58% of ImajevBench questions at 97.5% accuracy; at 99% it automates 40% at 100%.
- Apache-2.0 weights run under MLX on a Mac or PyTorch on one GPU; the card's sampled request answered in 1.15 s on a Mac Studio.
Watch out
Apache-2.0 vision-language adapter built on Qwen/Qwen3.5-4B. The leaderboard ranks are the card's own screenshots from late September 2026, and its demo set passes 130 of 145 combinations without the calibration file (118 with it); the card itself says to measure a few hundred of your own cases before choosing a threshold.
Questions about this entry
Short answers drawn from the linked project or page.
A 4B vision-language decision model built on Qwen/Qwen3.5-4B by Mohit Garg. It reads a photo against your own fields, or two photos in one request, and answers typed choice and noul questions with a probability on every option, including a trained can't tell.
The card shows the Apache-2.0 weights running under MLX on a Mac or PyTorch on one GPU, with an HTTP endpoint at /v1/systemone that accepts a request JSON plus an image file. The sampled request returned in 1.15 s on a Mac Studio with four option orders averaged and the calibration file applied.
Apache-2.0, as stated in the model card and its badges. The family also ships imajev-2b and imajev-9b on Hugging Face, and the card says photos and customer data never have to leave your network.
The card's screenshots show #1 of 91 on JevBench v1.4.2.2 with a 67.4 composite (scored 27 Sep 2026), #1 of 49 on Image JevBench v0.1.3 at 76.39, and #3 of 56 on DecisionBench (eng, v1) at 79.65, ahead of GLM-5.3 Flash, Jev 1.13, DeepSeek V4.1 Flash and GPT-5.6 Luna.
Every answer carries a probability for unknown and an abstained flag, and the card's automation table reports that at a 90% threshold the model decides 58% of ImajevBench questions automatically with 97.5% of those right; the rest are meant for a person.
Mohit Garg, who describes building decision systems for operations, support and back-office teams around refunds, returns, ticket routing, and document and photo checks; the card links his site at mohitgarg.ai, the code on GitHub, and a live demo Space.
No. It adopts TypeSafe Jev's request and response contract (POST /v1/systemone) and adds images, unknown_probability and abstained, but Jev is text-only and hosted while imajev is a separate small open model; the card notes Jev's 32k-token state limit exceeds imajev's 32 KB.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
Perplexity's 27B decision model rebuilt on Qwen3.8-27B with a joint request format that answers a whole request's questions in one sequence, reporting 68.09 on Decision Index 0.3 and 77.43 on vision.
Jev Models & Reimplementations#community#open-models#multimodal
Community
A decision model that scores its own clarifying questions by value of information: at AUC .799 against an oracle .797, it asks only when asking is expected to change the answer, then decides or hands off.
Jev Models & Reimplementations#community#open-models#local
Community
A small distilled sibling of Perplexity's pplx-decider-v1.2: a Qwen3.5-4B decider that reads a state plus typed choice, noul and score questions in one pass, keeping most of the 27B model's Decision Index.
Jev Models & Reimplementations#community#open-models#multimodal
Community
A Q8_0 GGUF of knowlet's Gemma-4-12B-Unified-System-One with a native System One decision runtime and an independently fitted GGUF temperature, shipped with an F16 mmproj for the model's native image and audio projections.
Jev Models & Reimplementations#community#open-models#quantized
Community
Fluid Inference's fixed-shape Core ML conversion of the classification path of Fastino's GLiNER2.5-Decide for Apple silicon: the DeBERTa-v3-large encoder and shared label classifier score up to four decision heads per call, in fp16, 8-bit or 6-bit packages.
Jev Models & Reimplementations#community#open-models#local
Community
OlaVerse's Apache-2.0 System One decision model from a stripped Qwen3.5-9B-Base: it returns calibrated choice, noul and score distributions for English and eight African languages in one pass, up to 65,536 tokens.
Jev Models & Reimplementations#community#open-models#multilingual
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.