Instructions to use TushaeBXN/matus-3b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TushaeBXN/matus-3b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="TushaeBXN/matus-3b",
	filename="matus-3b-Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use TushaeBXN/matus-3b with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf TushaeBXN/matus-3b:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf TushaeBXN/matus-3b:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf TushaeBXN/matus-3b:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf TushaeBXN/matus-3b:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf TushaeBXN/matus-3b:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf TushaeBXN/matus-3b:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf TushaeBXN/matus-3b:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf TushaeBXN/matus-3b:Q4_K_M

Use Docker

docker model run hf.co/TushaeBXN/matus-3b:Q4_K_M

LM Studio
Jan
Ollama
How to use TushaeBXN/matus-3b with Ollama:
```
ollama run hf.co/TushaeBXN/matus-3b:Q4_K_M
```

Unsloth Studio

How to use TushaeBXN/matus-3b with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for TushaeBXN/matus-3b to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for TushaeBXN/matus-3b to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for TushaeBXN/matus-3b to start chatting

How to use TushaeBXN/matus-3b with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf TushaeBXN/matus-3b:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "TushaeBXN/matus-3b:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use TushaeBXN/matus-3b with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf TushaeBXN/matus-3b:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default TushaeBXN/matus-3b:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use TushaeBXN/matus-3b with Docker Model Runner:
```
docker model run hf.co/TushaeBXN/matus-3b:Q4_K_M
```

Lemonade

How to use TushaeBXN/matus-3b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull TushaeBXN/matus-3b:Q4_K_M

Run and chat with the model

lemonade run user.matus-3b-Q4_K_M

List all available models

lemonade list

Matus 3B — K-12 Math AI Tutor

Built by Brian T. Thomas in collaboration with Dr. Raketa Ouedraogo-Thomas Independent ML/AI Developer | San Diego, California Full Sail University — B.S. Entertainment Business/ Model Updates coming soon!

What Matus Is

Matus is a fine-tuned Llama 3.2 3B model built for K-12 and early college math tutoring. It is the backbone of Project Matus — an open-source, locally-run AI tutoring system designed for students whose ways of knowing have historically been left out of math education.

Core design principles:

Never gives answers directly — scaffolds student thinking through questions
Recognizes valid alternative mathematical frameworks before correcting
Protects productive struggle — does not recalibrate downward because a student is frustrated
Culturally responsive — code-switching, family methods, and non-Western approaches are assets
Neurodivergence-aware — short responses and flat affect are not treated as disengagement
No data collection. No cloud. No subscriptions. Runs entirely on local hardware.

What It Runs On

Format: GGUF Q4_K_M quantization
Compatible with: llama.cpp, Ollama, LM Studio, any GGUF-compatible runtime
Minimum hardware: 8 GB RAM, any CPU (Intel, AMD, Apple Silicon)
Recommended: 16 GB RAM for comfortable performance
Latency: 30–90 seconds per response on CPU-only hardware. Faster on Apple Silicon or GPU.

How To Run It

One command with Ollama:

ollama run hf.co/TushaeBXN/matus-3b:Q4_K_M

One command with llama.cpp:

llama-server -hf TushaeBXN/matus-3b:Q4_K_M

With the full Project Matus system (auto-downloads model):

git clone https://github.com/TushaeBXN/project-matus.git
cd project-matus
pip install -r requirements.txt
./start.sh

K-12 Math Tutor (with simulated student profiles):

./boot_server.sh
python3 tutor/main.py --student james --role teacher

Training

Base model: Llama 3.2 3B Instruct (unsloth/Llama-3.2-3B-Instruct)
Method: LoRA fine-tuning (r=16, lora_alpha=32)
Dataset: 250 curated examples — identity data, conversational responses, K-12 and early college math tutoring scenarios across 9 domains
Training hardware: NVIDIA RTX A6000 (48GB VRAM) via RunPod
Framework: Unsloth + TRL SFTTrainer
Epochs: 3

Evaluation Results

Evaluated on 15 held-out math tutoring problems against a prompt-only baseline:

Metric	Baseline	Matus 3B	Target
Answer giveaway rate	0.0%	0.0%	<5%
Scaffolding quality	70.0%	72.7%	>70%
Conceptual accuracy	23.3%	32.6%	>70%
Ends with question	67%	86%	—

Behavior improvements over baseline: honor_struggle, affirm_partial, recognize_fallacy, counterexample, explain_composition all moved from 0% to 100%.

Full evaluation report: docs/preliminary_data.md

What It Won't Do

No data collection — nothing leaves your machine
No cloud dependency — runs fully offline after download
No answer giveaways — designed to scaffold, not solve
No biometric input — affect detection is text-based only
No diagnostic labeling — student memory stores behavioral observations, never deficit labels

Project

Part of Project Matus — an open-source K-12 math tutoring platform with:

Thought-token reasoning pipeline (internal reasoning hidden from student)
Affect detection (Tier 1/2/3 escalation) grounded in transformative SEL
Cross-session student memory (detects students who disengage across multiple sessions)
Epistemically just session logging (annotation-ready for researcher review)
5 simulated student profiles for testing and co-design

GitHub: github.com/TushaeBXN/project-matus

License

Apache 2.0 — free to use, modify, and distribute with attribution. Base model license: Llama 3.2 Community License (Meta).

Downloads last month: 10

GGUF

Model size

3B params

Architecture

llama

Hardware compatibility

4-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for TushaeBXN/matus-3b

Base model

meta-llama/Llama-3.2-3B-Instruct

Quantized

(461)

this model