Ministral-3-14B-Reasoning-2512-MLX-4bit

This is a 4-bit quantized MLX version of Ministral-3-14B-Reasoning-2512 for Apple Silicon Macs.

Known Limitations

Vision capabilities are NOT working in this MLX conversion. The model runs text-only inference successfully, but the Pixtral vision encoder does not properly process images. This appears to be a known issue with mlx-vlm's Mistral3/Pixtral support. Use this model for text-only tasks until mlx-vlm fixes Mistral3 vision support.

Model Details

Property	Value
Original Model	mistralai/Ministral-3-14B-Reasoning-2512
Parameters	14B (13.5B LLM + 0.4B Vision)
Quantization	4-bit (group size 64)
Size	~7.9 GB
Framework	MLX
Context Length	256K tokens
Vision Support	Not working (see above)

What Works

• Text generation: Full reasoning capabilities with [THINK] tags
• Multilingual: 11 languages supported
• Function calling: Native tool use support
• Performance: ~45-50 tokens/sec on Apple Silicon

What Doesn't Work

• Vision/Image understanding: The Pixtral vision encoder is included but does not properly process images due to mlx-vlm compatibility issues

Requirements

• macOS 15.0+ (Sequoia)
• Apple Silicon Mac (M1/M2/M3/M4)
• 16GB+ unified memory recommended
• Python 3.10+

Installation

pip install mlx-vlm

Usage (Text-Only)

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

# Load model
model, processor = load("hunterbown/Ministral-3-14B-Reasoning-2512-MLX-4bit")

# Text inference with reasoning
prompt = apply_chat_template(
    processor,
    config=model.config,
    prompt="Solve this step by step: What is 15% of 240?"
)
output = generate(model, processor, prompt, max_tokens=500)
print(output.text)

Performance

On Apple Silicon (M-series):

• Generation speed: ~45-50 tokens/sec
• Peak memory: ~8.5 GB
• Prompt processing: ~220 tokens/sec

Conversion Details

Converted using mlx-vlm:

python -m mlx_vlm.convert \
  --hf-path mistralai/Ministral-3-14B-Reasoning-2512 \
  --mlx-path ./ministral-3-14b-reasoning-4bit \
  -q --q-bits 4 --q-group-size 64

Alternatives for Vision

If you need vision capabilities, consider:

• GGUF versions with llama.cpp
• Wait for mlx-vlm to fix Mistral3 vision support

License

Apache 2.0 (same as original model)

Credits

• Original model by Mistral AI
• MLX conversion using mlx-vlm
• Quantized by @hunterbown