AgentJev
Open-weight 0.6B decision model on a Qwen3 backbone: state and typed questions in, a probability for every option out, nothing decoded.
Overview
AgentJev-0.6B is an open-weight, Jev-style System One model with weights on Hugging Face. It does not write. You give it unstructured state (a diff, stack trace, ticket thread or table) and questions you have already phrased, and one forward pass returns a probability for every option with zero decoded tokens. It is meant to sit where an agent loop needs a gate, a route or a score, leaving prose to a larger model.
It supports three primitives that each return the full distribution: Boolean (probability of true), Choice over 2 to 255 described options (value, top probability, margin) and Score over 2 to 10 ordered levels (argmax level and expected score). Question ids are not shown to the model, so meaning must live in the question and option text. The README hero reports 79.25% top-1 on 2,000 decisions with a 2,048-token context, and list summaries cite about 50 ms per forward pass.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Did the tests pass? / Is this command safe to run? | true · false |
| Choice | Which tool is next? | — |
| Score | Where on this ordered rubric does the state fall? | — |
How it works
- Serialize the state (objects become stable JSON)
- Phrase typed questions with descriptive options
- Run one forward pass
- Threshold the probability, margin or expected score in code
Strengths
- Runs locally with open weights
- No decoding, so no malformed output
- Full distributions with margin for thresholding
- Reported ~50 ms per pass
Caveats
- Accuracy figure is the author's own benchmark
- 2,048-token context limit
- Small model; quality on unfamiliar domains unknown
Getting started
Download weights from https://huggingface.co/aimeigaoshou/agent-jev and follow the README's Run it and HTTP sections.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- local model / HTTP server
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
AgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any…
wh000wh000/awesome-jev-liveAgentJev-0.6B - a fast 'System One' decision model for AI Agents: feed it any unstructured state (diffs, traces, logs) and structured questions, get calibrated probability distributions back in one ~50ms forward pass. Zero output-token decoding.
hellogumbo/awesome-jevA 0.6B decision model on a Qwen3 backbone with weights on Hugging Face: state in, a distribution over your options out, nothing decoded.
MrJev/awesome-jev
Frequently asked questions
How is it different from prompting a small LLM?
It decodes nothing; one forward pass returns probabilities over your options, so there is no output to parse.
Does it need a TypeSafe key?
No, it is an independent open-weight model you run yourself.
How accurate is it?
The README reports 79.25% top-1 on 2,000 decisions.
Related projects
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython