jevos-v4

jevos-v4

A local decision model that answers yes-or-no, multiple-choice and score questions about a text on a laptop CPU in about 28 to 130 ms, returning a probability for each answer. For developers who want a free local alternative to a cloud decision API.

Made by feder-cr on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: jevos-v4, published as feder-cr/jev, is a small model that makes decisions about a piece of text. You send a text and a question, which can be yes or no, multiple choice or a score, and you get back the probability of each answer. It runs on your own machine on the CPU, with no Python, and uses the same API format as TypeSafe Jev.

Who it suits: Developers who need quick classification decisions inside an app and want them local and free.

Industry | Who uses it | How they use it | Example
Customer support | Support teams | Route messages and spot billing issues | Is this a billing problem
E-commerce | Operations teams | Decide on refunds under a written policy | Should this customer get a refund
Games | Indie developers | Make fast in-game decisions | A Snake bot picking its next move
Software | Backend developers | Add a local classifier to a service | A triage step before a human reviews

You can skip it if you need the highest accuracy on hard tasks. The card says Jev's cloud API is more accurate on fraud and patent tasks, so test it on your own questions.

Where to use it:

• Your own machine, as a local server that uses only the CPU
• Windows, Linux and macOS on Apple silicon
• llama.cpp and tools built on it, through the GGUF files

Use case | Who | What they get back
Yes or no | Developers | The probability of yes
Multiple choice | Developers | The most probable option and a probability for each
Scores | Developers | The expected level, marked as early support

How to set it up:

• Python and pip, only for the hf download command
• The jev program for your system, taken from the project's GitHub release
• The model folder, downloaded next to it

1. Download the jev program for your system from the GitHub release.
2. Unpack it and download the model folder with the hf command.
3. Move into the folder and start the server.

tar -xzf jev-linux-x64.tar.gz                 # Windows: unzip jev-windows-x64.zip
hf download feder-cr/jev --include "model/*" --local-dir jev
cd jev
./jev serve                                   # Windows: jev.exe serve

How to use it: Send a POST request to the local server's systemone path with a text and your questions. One request can hold several questions about the same text. Each has a type, instructions and, for choices and scores, criteria.

{
  "model": "jev-latest",
  "state": {"item": "wireless mouse", "customer_message": "The box arrived empty. This is the second time!"},
  "questions": {
    "refund": {"type": "noul", "instructions": "Our policy refunds items reported missing within 30 days of delivery. Should this customer get a refund?"},
    "team": {"type": "choice", "instructions": "Which team should handle this message?",
             "criteria": {"billing": "payments, refunds", "shipping": "deliveries, missing parcels", "tech": "a product that does not work"}},
    "anger": {"type": "score", "instructions": "How angry is the customer?", "criteria": ["calm", "annoyed", "angry", "furious"]}
  }
}

Example: You send the message about a box that arrived empty with a billing question. The server replies with a probability for the answer. The card's own sample shows 0.94 for a duplicate charge.

{"model": "jevos-v4", "answers": {"billing": {"type": "noul", "noul": 0.94}}, "usage": {"input_tokens": 27, "output_tokens": 0}}

For newcomers and experts: If you are new, run the three setup commands and send the card's sample question. Check the answers against your own cases before you trust them.

If you know the field, the model has 882M parameters and ships as an OpenVINO INT8 folder plus 4-bit and 8-bit GGUF files, with an 8,192-token context. Choices and scores also return a confidence. When it is low, the card says to send the case to a person.

Your text plus questions
Local jev server on the CPU
Probability for every answer
Your app decides or escalates

Title: Median latency on a short request, in ms
jevos-v4: 28 ms
Laya: 129 ms
Jev cloud: 311 ms
Qwen3.5-4B: 3060 ms

Good to know:

• License: MIT
• Cost to run: free, on your own CPU, while the card says the cloud Jev charges per token
• Data it reaches: the texts you send to your own local server
• Not for: sums and counting, which belong in your own code
• Accuracy: lower than the cloud Jev on harder tasks, per the card

On Hugging Face, 13 November 2026

• Task: text classification
• Library: gguf
• Parameters: 882M
• Weights: about 441 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 1,475
• Likes: 22
• License: mit
• Languages: en
• Tags: gguf, openvino, llama.cpp, decision-model, yes-no, zero-shot-classification, local, on-device, text-classification, conversational

Links

• Model: https://huggingface.co/feder-cr/jev
• Files: https://huggingface.co/feder-cr/jev/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download feder-cr/jev

Then follow the usage in its model card. Nothing in this kit runs by itself.