Can I use it commercially?

No. The weights are CC BY-NC-4.0, forced by non-commercial training data even though the Qwen base is Apache-2.0; the Qwen license ships alongside. The base model is permissive but the checkpoint is what you deploy.

How does it handle option order?

It was built for it: options start at the same position, never attend to each other, and the head is position-free. The card reports 960 random shuffles of 3–12 options with no probability moving more than 1e-5 in fp32; bf16 adds about 0.001, which can only flip near-ties.

How do I install and run bev-decider?

Install the library with pip install bev-decider, or add the [serve] extra for the API. bev-decider serve --port 8008 exposes /v1/systemone, and the package downloads the single 0.97 GB model.safetensors checkpoint on first use.

What hardware does it need?

Very little: it is roughly 0.4B parameters, the first 20 of Qwen3-0.6B's 28 layers plus a small decision head, and runs on a laptop CPU or Apple Silicon. No GPU is required, though the default bf16 inference adds about 0.001 of rounding noise.

How well calibrated are its probabilities?

Expected calibration error is 0.065 on sysone-bench, so the card says its probabilities can be thresholded directly. The head was trained with cross-entropy on about 940K typed questions and needs no post-hoc temperature.

What input length and option limits apply?

States were trained up to 1,024 tokens; the defaults and benchmarks use a 2,048-token limit, and options are capped at 64 tokens each. The model runs a 2,048-token state against a 20-layer backbone.

Date and arithmetic comparisons are its least reliable area, along with stated exceptions to a general rule and 5-level sentiment, which is weaker than binary or categorical questions. It was trained mostly on English.

CommunityJev Models & Reimplementations

bev-decider 0.4B

A 0.4B Jev-compatible decision model with measured order invariance: 960 random option shuffles moved no probability by more than 1e-5, and bev-decider serve exposes /v1/systemone.

See highlights

Open huggingface.co

Published by: Community
Author: avbiswas
Last verified: Oct 2026

Tagscommunity open-models local scoring

Works forJev for Classification Jev for Scoring Jev for Routing

Used inJev for Software & DevTools

Languages: Python
Capabilities: Open sourceSelf-hostedRuns locally

Highlights

Order invariance measured, not asserted: 960 shuffles of 3–12 options moved no fp32 probability by more than 1e-5.
74.7 on its held-out typed decisions against Jev 1.13's 78.0, and ahead on score questions, 63.7 against 59.3.
On sysone-bench v2 it scores 70.8 against Jev's 90.7 but edges Kev and Laya on MNLI (62.5 against 48.3 and 55.8).
bev-decider serve exposes /v1/systemone on port 8008 from a single 0.97 GB checkpoint.

Quickstart

bash

pip install "bev-decider[serve]"
bev-decider serve --port 8008

curl -s localhost:8008/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "The customer asks for a refund after 40 days.",
  "questions": {"refund": {"type": "noul", "instructions": "Is this within policy?"}}
}'

Watch out

CC BY-NC-4.0 because of non-commercial training data, so no commercial use. State was trained up to 1,024 tokens, options cap at 64 tokens each, the held-out split shares the training distribution, and the card names date and arithmetic comparisons and 5-level sentiment as its weakest areas.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

Hacker News: Show HN: Ad Spend Guard gates budget before the auction, not the creative after.

Show HN: Gating ad budget with a decision modelRead the thread on Hacker News

X: Field Advisory sends marginal readings to an agronomist instead of a spray.

This post couldn't be loaded here — read it on X.https://x.com/agriwisehq/status/2781000000000000003

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 News

Related terms

Glossary definitions related to this entry.

More Jev resources like this

Gero-4B

A 4B scorer rebuilt from Qwen3-4B into a branching cross-encoder: one branch per option, no position bias, up to 256 options, and calibrated probabilities with no text to parse.