Small 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.
View on GitHub →Checked, not fully reviewed
Open 2B to 9B models that answer Jev-style questions about a photo and a record. Checked, not fully reviewed: it works on CPU, but takes minutes per request.
Good for
Watch out for
Tested Oct 27, 2026 at 318e8dad55c1 · python:3.11-slim in Docker on 2 CPU cores and 6 GB, CPU PyTorch, no GPU; the released imajev-2b adapter (Hub e8cb9fc6) on Qwen3.5-2B (15852e8), loaded in bfloat16 through a small wrapper, with its calibration file; the official typesafe-sdk 0.7.5 as a client
Checked, not fully reviewed. We ran its test suite and served the released 2B on two CPU cores. We sent it the README’s ticket example, its listing photo with two different records, a question the photo cannot answer, a set of bad requests and the official Python SDK. We did not run the 4B or the 9B, the Mac (MLX) path, a GPU, the browser playground, the benchmarks or the training. No Jev calls. The README’s leaderboard ranks and benchmark results are the author’s.
imajev fine-tunes Qwen3.5 with a LoRA and a small decision head. It serves them behind Jev’s
/v1/systemone, and accepts up to two photos. It does not call Jev. The serving path makes no
network calls, and it refuses image URLs, so only the model download leaves your machine.
It works, slowly. On CPU the server loads the model in float32. That ran out of our 6 GB, so we
used a wrapper to load it in bfloat16 instead. The ticket went to billing at 0.97, and its confidence
followed TypeSafe’s formula, scaled down by the unknown share. With the photo of beige shoes listed as
red, it named listing.color as the wrong field at 0.85. With the listing changed to beige, the colour
check went to 0.72, but the field question was unsure and picked the wrong answer, at 0.46. Asked for a
shoe size the photo does not show, it abstained, with 0.67 on unknown. Each request took between 34 seconds
and 3 minutes 22. Bad JSON, an unknown question type, nine questions and an eleven-level score each got a
422 with a plain message.
Two things to know. The official SDK worked for text. But it sends a photo as an assets upload,
which the server ignores: it answered from the text alone and gave no error. And it accepted a request
sent from another site’s origin. 639 tests passed and 8 failed. Six of those fail on Python 3.11, which the
README asks for, because one file needs Python 3.12.
Apache-2.0 for the code and adapters; the Qwen bases are Apache-2.0. 410 stars.
See how it compares with other tools in Best Jev tools, tested hands-on.
Review updated Oct 27, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.
NandhaKishorM/laya
Non-autoregressive decision engine over 100+ languages: three checkpoints and a router that detects the script and dispatches per request. Its benchmarks end with a limits section naming the datasets it does not generalise to and the headline figure that came from a training split.
jaredpalmer/kev
Jev-style decision models from 0.5B to 8B, built as LoRA adapters on Qwen and served behind a Jev-compatible /v1/systemone API.
TheoLeeCJ/SemIf-OpenJev
Jev-style decisions from a frozen 4B model on a single RTX 3090, with a browser demo. Formerly OpenJev.
New Jev releases, pricing changes, and the best new projects, once a week. No spam; unsubscribe anytime.
Powered by Buttondown. See our privacy policy.