🦬 Zubr Mini 1.9 3B — MLX

Repository: alekringtonnn-ai/zubr-mini-1.9-3b-mlx

---

📖 What This Model Is

zubr-mini-1.9-3b-mlx is the MLX version of the model alekringtonnn-ai/zubr-mini-1.9-3b, adapted to run on Apple Silicon (M1/M2/M3/M4) via Apple's MLX framework.

The base model, zubr-mini-1.9-3b, is a 3-billion-parameter language model fine-tuned from Mistral 3 3B (Ministral-3-3B / Ministral-3-3B-Instruct-2512 by Mistral AI).

Key Features

• Multimodal (Vision): the model supports image recognition and Image-Text-to-Text tasks (image + text → text).
• Multilingual: language support expanded from 26 to 48 languages (including Russian).
• Performance: improved vocabulary, high inference speed, and deep instruction tuning.
• Lightweight: only 3B parameters — the model runs fast locally even on entry-level Macs.
• License: Apache 2.0 — free to use, including for commercial purposes.

Model Tree

mistralai/Ministral-3-3B-Base-2512
        │ (quantization / instruct version)
mistralai/Ministral-3-3B-Instruct-2512
        │ (fine-tune by alekringtonnn-ai)
alekringtonnn-ai/zubr-mini-1.9-3b          ← GGUF version (Ollama, llama.cpp, vLLM)
        │ (conversion to MLX)
alekringtonnn-ai/zubr-mini-1.9-3b-mlx       ← MLX version (Apple Silicon) — THIS repository

---

📊 Versions (Quantizations) in the Repository

The MLX repository contains two weight files in .safetensors format with different quantization bit depths:

💡 How to choose: if in doubt, go with Q4. Q2 is a compromise for weaker hardware.

---

📥 How to Download

Requirements

• Mac with Apple Silicon (M1/M2/M3/M4) — MLX does not work on Intel Macs, Windows, or Linux with other vendors' GPUs.
• macOS 13.5+ (14+ recommended).
• Python 3.9+ (3.10+ recommended).
• Free disk space: ~4.3 GB for Q4 or ~3.2 GB for Q2.

Method 1: huggingface-cli (recommended)

# Install the CLI
pip install -U "huggingface_hub[cli]"

# Download the entire repository (both quants + README)
huggingface-cli download alekringtonnn-ai/zubr-mini-1.9-3b-mlx

# Or only a specific version:
huggingface-cli download alekringtonnn-ai/zubr-mini-1.9-3b-mlx \
  zubr-mini-1.9-3b-mlx-q4.safetensors

huggingface-cli download alekringtonnn-ai/zubr-mini-1.9-3b-mlx \
  zubr-mini-1.9-3b-mlx-q2.safetensors

By default, files are saved to ~/.cache/huggingface/hub/. The download path is printed to the console when finished.

Method 2: Direct links (browser / wget / curl)

# Q4 (recommended version)
curl -L -o zubr-mini-1.9-3b-mlx-q4.safetensors \
  "https://huggingface.co/alekringtonnn-ai/zubr-mini-1.9-3b-mlx/resolve/main/zubr-mini-1.9-3b-mlx-q4.safetensors"

# Q2 (compact version)
curl -L -o zubr-mini-1.9-3b-mlx-q2.safetensors \
  "https://huggingface.co/alekringtonnn-ai/zubr-mini-1.9-3b-mlx/resolve/main/zubr-mini-1.9-3b-mlx-q2.safetensors"

The same links also work in a browser: open the file page and click "Download", or append ?download=true to the URL.

Method 3: LM Studio (no command line)

1. Install LM Studio on your Mac.
2. In the search tab, type zubr-mini-1.9-3b-mlx.
3. Select the quantization you need (q4 or q2) and click Download.
4. The model will appear in your local models list — you can start chatting right away.

---

🚀 How to Run

Via mlx-lm (the official MLX library)

# Install
pip install -U mlx-lm

# Start chatting with the model (weights download automatically)
mlx_lm.generate \
  --model alekringtonnn-ai/zubr-mini-1.9-3b-mlx \
  --prompt "Hello! Tell me about yourself."

Via Python

from mlx_lm import load, generate

model, tokenizer = load("alekringtonnn-ai/zubr-mini-1.9-3b-mlx")

response = generate(
    model,
    tokenizer,
    prompt="Hello! Write a short poem about a bison.",
    max_tokens=512,
    verbose=True,
)
print(response)

Via LM Studio

1. Open LM Studio → the chat tab.
2. Select the downloaded zubr-mini-1.9-3b-mlx model from the dropdown list.
3. Set generation parameters (temperature, max tokens) and start chatting.

---

⚙️ Recommended Generation Parameters

---