Mica v0.1 4B

Mica v0.1 4B

Gives a probability for each allowed answer to a question about a given state, as yes/no, a choice or a score, without writing text. For developers who add quick decision checks to agents and apps.

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 is a small decision model. Its input is a situation, one question and a list of permitted answers. Its output is the likelihood of each answer. A question can be yes/no, a choice among 2 to 255 options, or a score with 2 to 10 levels. It reads the input a single time and writes no text, so one decision costs one prefill. It is trained on English and Korean.

Who it suits: It suits developers who want a fast, scored decision inside an agent or an app, instead of generated text to parse.

Industry | Who uses it | How they use it | Example
Agent development | Agent builders | Gate a risky action with a yes/no probability | Asking whether an agent may delete a staging database that nobody approved
Customer support | Support tool builders | Pick one category out of many options | Routing an incoming ticket to a queue
Content review | Moderation teams | Rate a text on a score scale of 2 to 10 levels | Scoring how polite a message is
Research | Evaluation researchers | Run the JevBench adapter against a local server | Checking the public hard tier on your own GPU

If you need a model that writes paragraphs or chats, you can skip it, since it never generates text.

Where to use it:

• Docker on a machine with an NVIDIA GPU
• Linux with CUDA 12 and Python 3.10 or newer, without Docker
• Clients that speak TypeSafe's System One request format, including ones written for Jev
• A local HTTP server on port 8010

Use case | Who | What they get back
Approval gate | Agent developers | A yes/no probability before an action runs
Option picking | App developers | A probability for each of up to 255 options
Score rating | Reviewers | A probability for each level of a 2 to 10 level scale

How to set it up:

• Docker with GPU support, or Linux with CUDA 12 and Python 3.10 or newer
• The project's code, Dockerfile and scripts, from its GitHub page
• Disk space for the BF16 GGUF, which is 9.7 GB, or a smaller quantized file

1. Get the project code from its GitHub page.
2. Build the Docker image.
3. Run the container with all GPUs and port 8010. The first start downloads the BF16 GGUF.
4. For a smaller download, pass the MICA_GGUF environment variable with the Q5_K_M file name.
5. Out of the box, the build covers CUDA architecture numbers 86, 89 and 90, and a CUDA_ARCH build argument changes them.

docker build -t mica-v0.1-4b .
docker run --gpus all -p 8010:8010 -v mica-weights:/weights mica-v0.1-4b

Without Docker, use these commands on Linux.

pip install -r requirements.txt
bash scripts/build_runtime.sh            # builds llama.cpp b11010 into ./runtime (CUDA_ARCH=86 by default)
hf download sky7350/Mica-v0.1-4B mica-v0.1-4b-BF16.gguf tokenizer.json tokenizer_config.json chat_template.jinja config.json \
   --revision ca36594cc2067c7252704f9f304cc10ef11c7c5c --local-dir weights
bash scripts/serve.sh

How to use it: With the server running on port 8010, send a POST request to its v1 systemone route. The body has a state, written as plain text, and a questions map. Each question has a type and instructions. The type noul means yes/no. Input longer than 8,192 tokens gets an HTTP 400 error and is not truncated.

{
  "state": "The user asked to delete the staging database. No approval has been given.",
  "questions": {"q": {"type": "noul", "instructions": "Should the agent delete it now?"}}}

Tool | What it does
noul | The question type for yes/no answers
scripts/build_runtime.sh | Builds llama.cpp into the runtime folder
scripts/serve.sh | Starts the local server
MICA_GGUF | Picks a smaller GGUF file when the container starts

Example: You send a state saying a user asked to remove the staging database and nobody approved it. The question asks whether the agent should delete it now. The reply holds a probability for each answer. The card does not show the numbers.

For newcomers and experts: A decision model gives you odds instead of sentences. You describe a situation, ask one question, and read how likely each answer is. A good first try is to start the Docker container and send the request above.

Under the hood, the base is Qwen3.5-4B. A LoRA of rank 16, already merged into the weights, covers each of its 32 layers. GGUF files come in six precisions, from Q4_0 up to BF16, with Q4_K_M, Q5_K_M, Q6_K and Q8_0 in between, and a merged BF16 safetensors copy is included. On an RTX 3090 with the BF16 GGUF and one request at a time, the card reports a median latency of 54 ms and a p95 of 552 ms.

State, question and allowed answers
One prefill pass
Probability for each answer
Returned by the local server

Good to know:

Title: JevBench hard tier accuracy on the public set
Mica: 69.5%
JEV 1.13: 74.3%
Qwen3.5-4B: 61.0%
JevK5 4B: 76.2%
Kev 4B: 52.4%
Nimble 9B: 23.8%
Laya: 10.5%

• Served by llama.cpp and driven by the JevBench runner, the public hard tier lands at 64.9, against 69.5 from the card's own PyTorch code. The two easier tiers do not change.
• On several other sets in the card, such as MMLU-Pro, Mica scores below JEV 1.13.
• The card says no JevBench item was used in training, and 77,732 training rows were checked against the 231 public items.
• License: Apache 2.0
• Cost to run: none named, though the card's setup needs an NVIDIA GPU
• Data it reaches: the text you send to your own local server, and the first Docker start downloads the weights
• Not for: writing text, and inputs longer than 8,192 tokens

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 4.2B
• Weights: about 8.4 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 3,436
• Likes: 52
• License: apache-2.0
• Languages: en, ko
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 2
• Fine-tunes: 1
• Tags: transformers, safetensors, gguf, qwen3_5_text, text-generation, decision-model, classification, judge, typesafe, jevbench, text-classification, conversational

Links

• Model: https://huggingface.co/sky7350/Mica-v0.1-4B
• Files: https://huggingface.co/sky7350/Mica-v0.1-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sky7350/Mica-v0.1-4B

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