Mica-v0.5-4B

Mica-v0.5-4B

A 4B classifier built on Qwen3.5-4B that reads an input and returns a probability for each allowed answer, for guardrails, routing, triage and agent gates. For developers who want yes/no or label calls without a large model.

Made by sky7350 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: Mica v0.5 4B is a compact classifier built on Qwen3.5-4B. For each question about your input, it rates every answer option you provide with a probability. You give it a log, trace, document or JSON of up to 32k tokens, plus the questions and allowed answers. Its output is only probabilities, so no text needs parsing and every result maps to an option you set.

Who it suits: Developers who call a large model only to get a yes or no or a label, and who want a small, quick classifier in its place.

Industry | Who uses it | How they use it | Example
AI engineering | Teams running guardrails | Check a request against allowed answers | Decide whether a message is safe to pass on
Retrieval | RAG developers | Check whether retrieved text answers the question | Score a retrieved passage as relevant or not
Agent systems | Builders of agents | Gate each step with a small choice question | Pick the next action from a list of options
Operations | Triage and routing teams | Route items by label | Label an incoming log or ticket

You can skip it if you need free-text answers, since it only returns probabilities for options you supply.

Where to use it:

• A single consumer GPU, with GGUF files for llama.cpp
• Python with the transformers library, using the safetensors files
• Guardrails, RAG checks, agent gates, routing, triage and evaluation

Use case | Who | What they get back
Guardrail check | Safety teams | A probability for each allowed verdict
Score a result | Evaluators | A level from a scale of 2 to 255 levels
Choose among options | Agent builders | A probability for each of up to 255 options

How to set it up: You need Python and a machine with one consumer GPU. The model comes as safetensors and GGUF, at about 9.3 GB in BF16 or about 2.3 GB at 4-bit by estimate. The card's quick start is cut off in the text used here, so follow it on the model page.

• Python and pip
• One consumer GPU, with disk space for the files
• The project's GitHub repository, for the request format and examples

1. Download the model files.
2. Choose the safetensors or the GGUF files for your setup.
3. Follow the quick start on the model page.

pip install -U huggingface_hub
hf download sky7350/Mica-v0.5-4B

How to use it: Each request bundles the material to judge, your questions and the answers you permit. The model reads the input once and returns a probability for every allowed answer. Set abstain_threshold if you want it to flag cases it is unsure about.

Setting | What it does
Allowed answers | Two to 255 options for each question
Input | Up to 32k tokens of text, JSON, logs or traces
abstain_threshold | Flags uncertain cases instead of guessing

Example: In the card's Minecraft demo, a player asks an in-game assistant for some food and a netherite ingot. Mica breaks the request into a chain of small choice questions, such as how many items and which one. It returns probabilities for every option, and the demo shows them beside each step.

For newcomers and experts: If you are new, think of it as a fast judge. You hand it a text and a fixed list of answers, and it tells you how likely each one is. A first small task is a yes or no check on a short document.

For experts, a single pass over the input yields a calibrated distribution and no generated tokens. The card says shuffling the option order still gives the same pick 97.3 percent of the time. The scores come from the author's own runs on the same items, and the card warns that its bar charts do not start at zero.

Input text or JSON
Questions and allowed answers
Mica reads the input once
Probability for each answer

Good to know:

Title: JDI 0.3 score from the author's own runs
PPLX Decider Flash: 60.1
Jev: 59.1
Mica v0.5: 58.8
CoCo Decision: 55.4
MS Decision-1: 54.7
Clef: 53.0
Kev: 39.1
Mica v0.1: 37.2

• Languages: English, Korean and Chinese, with more mentioned
• Output: probabilities only, never free text
• Benchmarks: figures are the author's own runs, so check them for your task
• License: Apache 2.0
• Cost to run: no API key or paid service named
• Data it reaches: downloads files from Hugging Face, and the input you give it stays in your own process
• Not for: free-text answers or writing

On Hugging Face, 13 November 2026

• Task: text classification
• Parameters: 4.7B
• Weights: about 9.3 GB in BF16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 6,375
• Likes: 47
• License: apache-2.0
• Languages: en, ko, zh
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 1
• Fine-tunes: 1
• Tags: safetensors, gguf, qwen3_5, decision-model, decision, judge, llm-as-a-judge, classifier, guardrails, agents, calibration, llama.cpp

Links

• Model: https://huggingface.co/sky7350/Mica-v0.5-4B
• Files: https://huggingface.co/sky7350/Mica-v0.5-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sky7350/Mica-v0.5-4B

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