MiniMax-M2.5-SWAN-4bit-MLX

Mixed-precision quantized version of MiniMaxAI/MiniMax-M2.5 using SWAN.

SWAN beats uniform 4-bit: -1.9% PPL, -1.7% size.

Metrics

Metric	Value
Size	118 GB
Average bits	3.77
Framework	MLX
WikiText-2 PPL	8.787 (mean)
Uniform 4-bit PPL	8.957

Rate-Distortion Curve

Quality vs size trade-off from MINT MCKP allocator. ★ = optimal knee point.

Budget	Size	Avg Bits	Loss
95 GB	95.0 GB	3.1	155.2783
114 GB	114.2 GB	4.0	22.3760	★
133 GB	133.4 GB	4.5	14.9543
153 GB	152.6 GB	5.2	10.2674
172 GB	171.8 GB	6.0	5.5744
191 GB	191.1 GB	6.8	3.1894
210 GB	210.3 GB	7.6	1.9911
230 GB	229.5 GB	8.1	1.0657
249 GB	247.8 GB	8.9	0.9171
268 GB	267.8 GB	9.7	0.8132
287 GB	286.8 GB	10.4	0.7152
306 GB	305.8 GB	11.2	0.6174
326 GB	324.7 GB	11.9	0.5197
345 GB	343.7 GB	12.7	0.4221
364 GB	363.8 GB	13.5	0.3191
383 GB	382.8 GB	14.3	0.2216
402 GB	401.7 GB	15.0	0.1242
422 GB	420.7 GB	15.8	0.0269
441 GB	426.0 GB	16.0	0.0000
460 GB	426.0 GB	16.0	0.0000

Generated by MINT rate-distortion optimization.

Usage

from mlx_lm import load, generate

model, tokenizer = load("baa-ai/MiniMax-M2.5-SWAN-4bit-MLX")
response = generate(model, tokenizer, prompt="Hello!", max_tokens=256)
print(response)

About SWAN

SWAN uses data-free per-tensor sensitivity analysis with composite scoring to allocate bit-widths across model layers.

• Paper | Code | Models

---

Quantized by baa.ai