JEV-CPU

JEV-CPU

Reads a decision from a small model's option probabilities on an ordinary laptop CPU, with no GPU and a browser interface. For developers who want quick routing and triage choices inside agents.

Made by Meanblock 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: JEV-CPU makes decisions without writing an answer. You describe a situation, add a criterion and list the typed options, and it returns a probability for each option, read from the logits of the small Qwen3-0.6B model in about a second. It adapts the SemIf project to run on a plain CPU and adds a web interface. It is an independent project, not endorsed by SemIf's author.

Who it suits: Developers who build agents and need cheap, fast choices such as route this, retry that, or does the evidence support a claim, and who have no GPU.

Industry | Who uses it | How they use it | Example
Customer support | Support engineers | Classify incoming messages and rank their priority | A ticket marked urgent or routine
Content moderation | Trust and safety teams | Triage content against written criteria | A post sorted into review or allow
Software | Engineering leads | Triage code reviews and rate incident severity | A change tagged as low or high risk
Finance and compliance | Operations teams | Screen items against stated rules | A request flagged for a manual check

You can skip it if you need free-form written answers, since it only scores the options you declare.

Where to use it: It runs on a laptop CPU through Python and transformers, and it ships a browser interface.

• A laptop or desktop CPU on Windows, macOS or Linux
• Python with the transformers library
• The web UI that comes with the project

Use case | Who | What they get back
Routing | Agent developers | A probability for each destination
Content triage | Moderation teams | A probability for each decision option
Severity rating | Operations teams | A probability for each level

How to set it up: You need Python with pip, a CPU and about 2.4 GB for the float32 Qwen3-0.6B weights. The card has its own quick start section for running it.

• Python with pip
• A CPU, with no GPU required
• About 2.4 GB for the model weights in float32

1. Install the Hugging Face download tool.
2. Download the project files to your computer.
3. Follow the quick start section of the model card to launch the engine and the web UI.

pip install -U huggingface_hub
hf download Meanblock/JEV-CPU

How to use it: In the web UI you type the state, add your criteria and declare the typed options. The engine then returns the decision in about a second, with no text generated. Under the hood it takes the logits at the last position of one forward pass.

logits = model(**inputs, use_cache=False).logits[:, -1, :]

Tool | What it does
Web UI | Takes the state, the criteria and the options
CPU engine | Loads Qwen3-0.6B in float32 and scores the options
Option logits | Turn into the probability shown for each option

Example: Type a customer message as the state, add the criterion of how urgent it is, and declare options such as urgent and routine. The page shows a probability for each option, so you can see how sure the model is.

For newcomers and experts: Most language models write a sentence and you read the answer from it. This one skips the writing and reads which option the model leans toward, as numbers. A first small try is to type a short message in the web UI and give it two options.

For experts, every option gets one uppercase letter, and the prompt asks for only that letter with thinking switched off. The engine checks that each letter is a single token that survives a decode round trip, then runs one forward pass and applies a softmax over those letter slots. The trade is model size for hardware: a 0.6B model on a CPU is accessible, and the upstream project targets a 4B BF16 model on CUDA.

You type a state and add criteria
Each option gets a single-token letter
One forward pass runs on the CPU
Logits at the last position are read
A softmax gives a probability per option

Good to know: The card shows one live demo across eight domains and gives a decision time of about one second.

• Independent: The project is not affiliated with SemIf's author, TypeSafe or Jev.
• No generation: It never samples an answer, so there is no JSON to repair.
• Same scoring: It reuses the original SemIf scoring code and swaps only the model loader.

• License: MIT.
• Cost to run: none named, the source lists no API key, account or paid service.
• Data it reaches: the state and criteria you type in, on your own machine, plus the model files you download.
• Not for: generating written answers, since it reads option probabilities and writes no text.

On Hugging Face, 13 November 2026

• Task: zero shot classification
• Library: transformers
• Downloads in the last 30 days: 0
• Likes: 67
• License: mit
• Based on: Qwen/Qwen3-0.6B
• Tags: transformers, semantic-if, decision, semif, zero-shot-classification

Links

• Model: https://huggingface.co/Meanblock/JEV-CPU
• Files: https://huggingface.co/Meanblock/JEV-CPU/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Meanblock/JEV-CPU

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