imajev-4b is an Apache-2.0 LoRA on Qwen3.5-4B that reads up to two photos.

imajev-4b is the 4B checkpoint in Mohit Garg's mohit67890/imajev. The dated write-up, including the September boards, is the Imajev story. What a TypeSafe Jev call returns is on what Jev is.

What is imajev-4b?

imajev-4b is a LoRA on Qwen3.5-4B with a 256-code readout and a frozen vision tower. The LICENSE file is Apache-2.0.

The model card names base revision 851bf6e8, LoRA rank 64 and alpha 128 on the language layers, and a bias-free readout of 256 by 2560. adapter_model.safetensors is listed at 487.6 MB.

The card's request limits are 0 to 2 images, state up to 32 KB, 1 to 8 questions, and at most 4,096 tokens. It says longer requests are refused, and that the checkpoint is English only. TypeSafe's models page, opened October 1, prints 32k tokens for state plus the longest question on jev-1.13.0.

The README calls this size the default. The 2B and 9B weights are the previous adapters. On ImajevBench the README prints 83.9% for the 4B (234 of 279), 82.1% for the 9B, and a paired test of p = 0.57. Text answers use Choice, Score, and Noul, plus unknown_probability and an abstained flag. We did not run the server.

Where are the imajev-4b weights?

The weights are mohit67890/imajev-4b on Hugging Face. The code and the quickstart are in mohit67890/imajev.

The README links a live demo at the Hugging Face space mohit67890/imajev, and a site at mohit67890.github.io/imajev. We did not open the space.

How do you run imajev-4b?

The README quickstart clones the repo, installs the serve extra, downloads a pinned Qwen3.5-4B, and downloads mohit67890/imajev-4b into adapters/imajev-4b.

Apple silicon uses the mlx extra. Other machines use the torch extra. The server is scripts/playground/server.py, with --model-name imajev-4b, the mlx adapter, calibration.json, and port 8765. The endpoint is POST /v1/systemone. A sample listing photo in that README came back as model imajev-4b, choice listing.color, confidence 0.919, and total_ms 1152.9, on a Mac Studio with four option orders and the calibration file. The same README's size table prints 350 ms as shipped, and 96 ms raw, for a JevBench hard item on one H100. We did not run either measurement.

What did imajev-4b score on JevBench?

Opened October 8, 2026, the open-weights capability list, headline release v1.6.1 and revision v1.7.21, puts imajev-4b at 61.9, capability rank 31, official composite 47. Intelligence is 34.6, Calibration is 89.2, estimated cost is $0.017 per 1,000 decisions, and p50 is 0.29 s. On October 7, revision v1.7.12, the same card was capability rank 17 and official composite 26. We did not rerun the suite.

JevBench v1.5.6, the page we opened on October 4, 2026, lists imajev-4b at 70.4, official rank 10 of 110. The row is still tagged roster addendum A2.

The axis cells are 53.5, 88.1, 91.1, and 63.3. Estimated cost is $0.017 per 1,000 decisions. Cygnet, Winnow-12B Q8, and Jev 1.13.0 stay 73.7, 73.2, and 72.1, and the page still calls Cygnet and Winnow joint leaders. The detail row is imajev-4b on that board. The whole ranking is the JevBench story, and the separate explainer is Jev benchmark.

70.4 is the v1.5 point estimate. 67.37 is the v1.4.2.2 composite, changelog date September 27, rank 1 on that older scorer, with Jev still at 63.29. The capability view we opened September 28 lists 66.3 and official rank 1 on that board.

The page opened September 29 would place the addendum sixth at 70.4. The page opened September 30 numbers the same 70.4 as official rank 9. The author's public-hard line, four option orders plus the calibration file, is 72.1%. We did not rerun the boards.

On October 7, 2026 the same ranking URL says headline release v1.6.1 and board revision v1.7.12. The open-weights capability list puts imajev-4b at rank 17 and 61.9, Intelligence 34.6 and Calibration 89.2, estimated $0.017 per 1,000 decisions, p50 0.29 s. The rank line says official composite 26. 70.4 stays the v1.5.6 point estimate. We did not rerun v1.6.1.

What did imajev-4b score on Image JevBench?

Opened October 8, 2026, Image JevBench headlines JevImageBench v0.3.0. Imajev-4B leads the Jev-class systems at capability 68.0, rank 1, official composite rank 2 at 63.7 among 52 ranked systems. Intelligence on the capability card is 49.9, Calibration is 86.1, and the cost is $0.038 per 1,000 decisions, p50 0.37 s. We did not rerun the image board.

Image JevBench v0.1.5, opened October 4, lists imajev-4b at 76.4, official rank 2 of 50.

The compare table prints 76.39, with Intelligence 73.8, Calibration 90.5, Speed 87.6, and Cost 61.2, about $0.020 per 1,000. Wity-1 is rank 1 at 80.2, and that row says the server build ID was not recorded and the score may change after verification. Jev-Omni is 73.1 at rank 3. The setting line says PyTorch, one rotation, --fast, --merge-lora, and calibration.json. The author's September 28 screenshot of v0.1.3 prints 76.39 as rank 1 of 49. The capability view on v0.1.5 lists imajev-4b third at 82.1. Wity-1 leads that view at 85.7. We did not rerun it.

What else did the author publish for imajev-4b?

ImajevBench after the September 26 phase-3 adapter is 83.9%, up from 82.4%, on 279 questions. Unknown abstentions are 18 of 21, up from 14 of 21.

At a 90% bar the README automates 58% of that set and is right on 97.5% of those. At 99% it automates 40% and prints 100% right. At 80% it automates 63% and is right on 94.9%. Scenario checks are 129 of 145 without the calibration file and 113 of 145 with it. The top answer stays put, and confidence drops. A DecisionBench screenshot in the README prints imajev-4b at 79.65, third of 56, with jev-1.13 at 71.90. DecisionBench ECE in the September 26 note moves from 0.024 to 0.069. We did not load the live Gradio board. Measurements already on the desk, each with its caveat, are on Evals.