metadata
language:
- en
license: mit
pipeline_tag: text-generation
tags:
- mlx
- mixture-of-experts
- moe
- pruning
- reap
- minimax
- 4bit
- quantized
- apple-silicon
library_name: mlx
base_model: Akicou/MiniMax-M2-5-REAP-29
MiniMax-M2.5 REAP-29 — MLX 4-bit
MLX 4-bit quantized version of Akicou/MiniMax-M2-5-REAP-29 for efficient local inference on Apple Silicon.
- Quantization: 4-bit (group size 64, affine mode; router gates at 8-bit)
- Architecture: MiniMax M2.5 MoE — 62 layers, 180 experts (REAP-pruned from 256), 8 active per token
- Context: 196K tokens
- Size: ~85 GB
- Pruning: 29% of experts removed via REAP (Router Expert Activation Pruning)
Usage
from mlx_lm import load, generate
model, tokenizer = load("shieldstackllc/MiniMax-M2.5-REAP-29-mlx-4bit")
response = generate(model, tokenizer, prompt="Hello!", verbose=True)
Or with vMLX for native macOS inference.
About
MiniMax-M2.5 is a large Mixture-of-Experts language model by MiniMax AI. This variant was pruned to 29% fewer experts by Akicou using REAP (Router Expert Activation Pruning), reducing model size and memory footprint while maintaining strong performance. MLX quantization by vMLX.
Also Available
- MiniMax-M2.5-REAP-39 MLX 4-bit (~73 GB) — 39% pruned variant
- MiniMax-M2.5-REAP-39 MLX 8-bit (~138 GB) — 39% pruned variant
Made for vMLX
This model was converted and optimized for vMLX — a free, open source macOS native MLX inference engine for Apple Silicon. Download vMLX to run this model locally with zero configuration.
Credits
- Base model: MiniMaxAI/MiniMax-M2.5 by MiniMax AI
- REAP pruning: Akicou/MiniMax-M2-5-REAP-29 by Akicou
- MLX conversion: vMLX — Run AI locally on Mac. No compromises.
Contact
For questions, issues, or collaboration: admin@vmlx.net