jevos
Open-source Jev-compatible decision model served by one native binary on a laptop CPU, returning probabilities in 28 to 130 ms.
Overview
jevos is an MIT-licensed, Jev-compatible alternative that runs entirely on your own machine. A single native binary for Windows, Linux or macOS, with no Python, serves POST /v1/systemone in TypeSafe's wire format, so code written for Jev's SDK works unchanged. The model is published on Hugging Face and as GGUF for llama.cpp, and the current release ships an OpenVINO int8 build. Earlier versions, as described in some list entries, were a 1B MiniCPM-derived model answering only yes/no questions.
The current README says jevos-v4 answers Noul, Choice and Score questions (Score still early) with a probability for each answer, with an 8,192-token context. It reports median latency of 28 ms on short and 130 ms on long requests on a 16-thread Intel Core Ultra 7 laptop CPU, against about 311 to 314 ms for hosted Jev, and lower accuracy than hosted Jev on all six published task sets (for example 0.76 against 0.91 on a rental-policy set). It suits developers who want free, offline decisions and can accept a quality gap.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this a billing problem? (any yes/no question) | — |
| Choice | Which of the given options fits the text? | — |
| Score | Where on the given scale does the text fall? (early support) | — |
How it works
- Download the binary and model from the release or Hugging Face
- Run ./jev serve to start the local server on port 8017
- POST state and typed questions to /v1/systemone
- Read per-answer probabilities and apply thresholds in your code
Strengths
- Free and offline once downloaded
- Runs on CPU with no GPU or Python
- Drop-in for Jev's wire format
- Benchmarks published against hosted Jev and Laya
Caveats
- Lower accuracy than hosted Jev on every published task set
- Score support is early
- Five of six benchmark sets helped choose the released checkpoint
Getting started
Download the release archive and jevos-v4-openvino-int8.zip, unzip the model into jev/model, then run ./jev serve and POST to http://127.0.0.1:8017/v1/systemone.At a glance
- Type
- Service
- Status
- beta
- Runtime
- native binary server (CPU), GGUF for llama.cpp
- Has code
- yes
- License
- MIT
- Stars
- 1,312
Use cases and integrations
How curated lists describe it
Open alternative: MIT-licensed 1B model (MiniCPM5 cut to 17 layers with a one-logit head, GGUF q4_k_m at 619 MB) that answers only `Noul` yes/no questions on Jev's own `/v1/systemone` wire format, running CPU-only via llama.cpp at 54 ms short / 220 ms long on a laptop Core Ultra 7 255H against Jev's hosted 344/345 ms; `Choice` and `Score` return 422, and it scores 0.815 accuracy on 2,000 unseen policy yes/no questions against Jev's 0.927.
yibie/awesome-jev1B decision model for CPU-only laptops: MiniCPM5 cut to 17 layers with a one-logit head, GGUF q4_k_m at 619 MB, running on llama.cpp with no GPU, about 54 ms per short request. Speaks Jev's `/v1/systemone` wire format for Noul (yes/no) questions only; Choice and Score return 422.
AnotiaWang/awesome-decision-modelsjevos is an open-source, Jev-compatible alternative to TypeSafe's Jev for yes/no decisions that runs entirely on a laptop CPU. Send text plus a yes/no question over HTTP and get back P(yes) from a single forward pass of a 1B model — no text generation, no GPU required.
logicrw/awesome-jev-projectsjevos: an open-source alternative to Jev and Microsoft-Decision-1. Yes/no decisions, choices and scores with probabilities, on your laptop's CPU.
wh000wh000/awesome-jev-liveMIT decision model served by one native binary (Windows, Linux, macOS, no Python) behind Jev's `/v1/systemone` wire format on a laptop CPU, also published as GGUF for llama.cpp; answers Noul, Choice and Score questions (Score still early) with a probability for each answer while code keeps the thresholds; its README reports 28 ms on a short request and 130 ms on a long one on a 16-thread laptop CPU against 311-314 ms for hosted Jev, and lower accuracy than hosted Jev on all six of its published task sets (for example 0.76 against 0.91 on a rental-policy set).
cobanov/awesome-jevServes `/v1/systemone` and `/v1/models` for yes/no questions only — other types are refused — from a GGUF model through llama.cpp on CPU, offline once the file is downloaded.
MrJev/awesome-jev
Frequently asked questions
Can I use my Jev client with it?
Yes, it serves /v1/systemone in Jev's wire format so SDK code works unchanged.
How does it compare with hosted Jev?
It is faster on a laptop CPU per the README, but less accurate on every published task set.
Does it support Choice and Score?
Version 4 supports both, with Score still early; earlier versions answered yes/no only.
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