jevalt
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.
Overview
JevAlt provides three small decision models, Deem-4B (English), Karar-4B (Turkish) and Wahler-4B (German), fine-tuned from Intern-Decision-4B. Given a situation, a question and answer options, each chooses an option and returns a probability for every one, for tasks such as routing support tickets or checking a written policy. The Q4_K_M GGUF builds run on a CPU in about 3 GB of RAM with an optional GPU, so decisions can be inspected and tested on your own data.
A bundled server implements Jev's POST /v1/systemone API, so existing Jev clients can call it; documentation covers question types, optional reasoning and an unknown answer. On held-out decisions from the project's own data pipeline, the README reports 94.7% (English), 96.8% (Turkish) and 92.0% (German) accuracy, each about ten points above the same-size Kev-4B baseline, while noting these tests favour JevAlt. Calibration is fitted on project data and should be refitted on yours with the companion JevOss tooling. A Hugging Face Space offers a hosted example.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this ticket? (example) | billing · technical · sales |
| Other | Any Jev-style Choice, Score or Noul question, with an optional unknown answer | — |
How it works
- pip install and run jevalt serve (downloads a model)
- Send Jev-shaped requests to /v1/systemone
- Model returns choices with probabilities
- Refit calibration on your own data if needed
Strengths
- Runs on CPU in about 3 GB RAM
- Jev API compatible
- Language-specific models for Turkish and German
- README reports roughly ten points over Kev-4B per language
Caveats
- Evaluations come from the project's own pipeline and favour it
- Calibration fitted to project data
Getting started
pip install "jevalt[serve,gguf] @ git+https://github.com/mertkayacs/jevalt" && jevalt serveAt a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- Local server (GGUF)
- Has code
- yes
Use cases and integrations
How curated lists describe it
Three small 4B language models for local decisions in English, Turkish and…
wh000wh000/awesome-jev-liveOpen alternative to Jev, Kev and Laya: three 4B decision models (English, Turkish, German) that answer typed questions with calibrated probabilities. Jev API compatible, runs on a CPU in about 3 GB of RAM.
hellogumbo/awesome-jev
Frequently asked questions
Can I use existing Jev clients?
Yes, the server implements POST /v1/systemone.
Which languages are covered?
English, Turkish and German, one 4B model each.
How trustworthy are the reported accuracies?
The README notes the tests come from its own data pipeline and favour JevAlt.
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++Japanese System One decision model on ModernBERT-ja (314.6M) returning typed choice, score and bool distributions with no generation.
ProjectbetaPython