baa-ai/MiniMax-M2.5-SWAN-4bit

MiniMax-M2.5 — SWAN Mixed-Precision (4-bit avg)

This is MiniMax-M2.5 quantized using SWAN (Statistical Weight Analysis for N-bit allocation) — a data-free per-tensor mixed-precision quantization method for MLX on Apple Silicon.

Key Features

• Data-free quantization: No calibration dataset required — uses weight statistics only
• Per-tensor bit allocation: Each tensor gets 2, 4, 8, or 16-bit based on sensitivity analysis
• MLX native: Ready for inference on Apple Silicon via mlx_lm

Results

Metric	SWAN (this model)	Uniform 4-bit	SWAN vs Uniform
PPL (WikiText-2, mean)	8.787	8.957	-1.9%
PPL (WikiText-2, median)	9.169	9.399	-2.4%
PPL (WikiText-2, trimmed)	8.748	8.924	-2.0%
Model size	118 GB	120 GB	-1.7%
Peak memory	121 GB	123 GB	-1.6%

Evaluation config: WikiText-2 test split, sequence length 2048, 256 samples, seed 42.

Usage

pip install mlx-lm

# Generate text
python -m mlx_lm.generate \
    --model baa-ai/MiniMax-M2.5-SWAN-4bit \
    --prompt "Hello, how are you?"

# Interactive chat
python -m mlx_lm.chat --model baa-ai/MiniMax-M2.5-SWAN-4bit

Quantization Details

Setting	Value
Source model	MiniMaxAI/MiniMax-M2.5 (FP8, 229B params, 10B active)
Method	SWAN v4 (adaptive normalization + optimized thresholds)
Normalization	Adaptive (percentile-based, optimal for MoE)
Thresholds	t2=0.20, t8=0.75, t16=0.95 (grid-search optimized)
Average bits	3.77 bpw

Bit Distribution

Precision	Parameters	Percentage
2-bit (group_size=32)	90.7B	39.7%
4-bit (group_size=128)	115.6B	50.5%
8-bit (kept at FP8)	17.6B	7.7%
16-bit (protected)	4.8B	2.1%

Hardware Requirements

• Apple Silicon with at least 128 GB unified memory (192 GB recommended)
• Peak memory during inference: ~121 GB

About SWAN

SWAN computes four sensitivity metrics per tensor:

• SVD spectral concentration
• Excess kurtosis
• Output noise amplification
• Reconstruction error proxy (NRMSE)

These are combined into a composite score that drives automatic bit-width allocation — without any calibration data.

Paper: SWAN: Data-Free Mixed-Precision Quantization for LLMs via Multi-Metric Sensitivity Analysis (Black Sheep AI Research, 2026)