Open typed-decision models that return probabilities over candidates directly, without generating or parsing an answer, with a project site and reports.
View on GitHub →Checked, not fully reviewed
Open 2B to 27B decision models with TypeSafe's confidence formulas and careful errors. Checked, not fully reviewed: correct on CPU, but minutes per request.
Good for
Watch out for
Tested Oct 13, 2026 at bd4118882f73 · The repository's own CPU Docker image (open-jev:2b-cpu) with the released 2B checkpoint (Hub 0c7aa498) on Qwen3.5-2B (15852e8), limited to 2 CPU cores and 10 GB, no GPU; the official typesafe-sdk 0.7.4 as a client
Checked, not fully reviewed. We built the repository’s own CPU Docker image with the released 2B checkpoint, sent it the README’s example, the official SDK and a set of bad requests, and checked every confidence it returned. We had planned to run the public JevBench items as well. At three minutes a request on two cores that would have taken most of a day, so we did not. No Jev calls. The README’s comparisons with Jev and its 27B results are the author’s own.
Open-Jev trains a LoRA and a small decision head on Qwen3.5 (2B, 9B and a 27B v1.1), and serves them
behind Jev’s /v1/systemone. The 2B download is the adapter and head; the loader fetches the pinned
Qwen base itself. The Docker build bakes both in and checks they match, so the container needs no
network once built.
It gets the details right. For the README’s billing example, a choice at 0.728 / 0.041 / 0.231
came back with confidence: 0.592, and a score at 0.127 / 0.472 / 0.401 with 0.208. Both are
TypeSafe’s formulas exactly, and the code cites them. Malformed JSON, an unknown question type, a
missing questions and a twelve-level score each got a 422 with a plain message. A POST from a
different origin got a 403, so a web page cannot use the local server, and unknown paths get a
404. /v1/models lists jev-latest as an alias, and the official Python SDK worked unchanged.
It is slow without a GPU. The CPU image runs the plain PyTorch path. The fast kernels are GPU-only, the server answers one request at a time, and prefix caching is off by default for accuracy. A three-question request took 180 to 190 seconds on two cores, and one yes/no question took 22. The README says CPU is “substantially slower” and suggests short requests, and it is right to.
MIT for the code; the Qwen bases and the 2B checkpoint are Apache-2.0. 410 stars. For a smaller open model in Jev’s format that answers in seconds on a CPU through llama.cpp, see StartLux-Decision.
See how it compares with other tools in Best Jev tools, tested hands-on.
Review updated Oct 13, 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.