Kev-4B

Kev-4B

Reads a document and typed questions, then returns a calibrated probability for each answer option in one pass, without writing text. A Qwen3.5-4B based decision model for developers who route, triage or check documents.

Made by jaredpalmer 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: Kev-4B is a decision model, not a chat model. You give it a document and a few typed questions, and it scores the options for each question in one pass. The probabilities are calibrated, so you can set a threshold and send unsure cases to a person.

Who it suits: Developers who sort, route or vet documents in software and want a score they can set a threshold on. A single mid-range GPU is enough to host it. The files carry the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Route incoming messages by topic and urgency | Sending a double-charge complaint to billing
Operations | Workflow builders | Automate confident cases and queue the rest for review | Auto-approving a request only above a set confidence
Compliance | Policy teams | Check documents against stated rules or criteria | Checking whether an application meets eligibility rules
AI engineering | Developers | Judge a proposed answer against criteria | Scoring a draft reply as acceptable or not

Anyone who wants text written, chat or summaries can skip it, because the model only scores options you supply.

Where to use it:

• A CUDA GPU, with about 16 GB free, or an Apple Silicon Mac that runs it through MLX
• Python programs that call the TypeSafe SDK against your own server
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Document classification | Developers | A probability for each named category
Routing and triage | Operations teams | The most likely destination and a confidence value
Policy checks | Compliance teams | The chance that a yes or no question is yes

How to set it up:

• Python and the huggingface_hub package, to download the files
• The Kev repository from GitHub, plus the uv tool
• A CUDA GPU with about 16 GB free, or an Apple Silicon Mac

1. Download the model files with the commands below.
2. Clone the Kev repository from GitHub and install its serve extras with uv.
3. Start the server with one of the two serve commands in the next section.
4. Point the TypeSafe SDK at your server.

pip install -U huggingface_hub
hf download jaredpalmer/kev-4b

How to use it: The card serves the model with the Kev repository on port 8008. Your code then sends a state and questions through the TypeSafe SDK. The first command below follows the main branch, and the second pins the same weights.

uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8008           # Kev 1.0 (this card)
uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b@v1.0 --port 8008      # the same weights, pinned

Tool | What it does
choice | Scores 1 to 255 named options and picks the most likely one
score | Scores 1 to 255 ordered levels and gives an expected level
noul | Answers a yes or no question with the chance of yes
kev.train | Fine-tunes from the released weights on your own labels

Example: You send the text of a support message about a double charge, plus a choice question with the options billing, shipping and other. You get a probability for each option, so a low top score can go to a person.

For newcomers and experts: A classifier that also tells you how sure it is. A good first step is to start the server and ask one yes or no question about a short text.

Under the hood, the model is a LoRA adapter of rank 16 and a pointer head on a frozen Qwen3.5-4B-Base. Each option's closing token is scored against the question's final token, and a softmax gives the probabilities. One temperature, 2.41, calibrates them. Questions run as separate rows, so they cannot sway each other.

Send a document and typed questions
Compute the document state once
Score each question in its own row
Pointer head gives probabilities
Act on confident cases, queue the rest

Good to know:

• License: Apache 2.0
• Cost to run: none named, but you need a GPU or an Apple Silicon Mac
• Data it reaches: the download fetches files from Hugging Face, and your documents go to the server you start
• Language: English only
• Not for: writing text, chat or summaries, and fully automated decisions with legal, medical or financial effects on people
• The validated context is 8,192 tokens, and longer documents may do worse

Title: Context lengths in tokens
Served states, up to: 65536 tokens
Validated length: 8192 tokens
Longest training state: 7552 tokens

On Hugging Face, 13 November 2026

• Task: text classification
• Library: peft
• Formats: Safetensors
• Downloads in the last 30 days: 12,499
• Likes: 175
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-4B-Base
• Quantized versions: 14
• Fine-tunes: 2
• Tags: peft, safetensors, decision-model, calibration, lora, multiple-choice, typesafe, qwen3.5, text-classification, model-index

Links

• Model: https://huggingface.co/jaredpalmer/kev-4b
• Files: https://huggingface.co/jaredpalmer/kev-4b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download jaredpalmer/kev-4b

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