How much VRAM does MiniMax-M2.5 need?

Q4_K_M is exactly 138,342,384,064 bytes (128.84 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is MiniMax-M2.5's KV cache?

7.75 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

Is MiniMax-M2.5 a mixture-of-experts model?

Yes — 256 experts with 8 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

MiniMaxAI · text · mixture of experts

MiniMax-M2.5

MiniMaxAI/MiniMax-M2.5

MiniMax-M2.5 at Q4_K_M is exactly 138,342,384,064 bytes (128.84 GiB / 138.34 GB) — an effective 4.839 bits per weight, not the nominal 4. Its KV cache at 32K is 7.75 GiB.

From the file· summed from 4 file(s)From the file· KV per layer

Parameters

229B

total, not active

Architecture

minimax-m2

62 layers

Context

196,608

native (config.json)

License

other

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
I1-IQ1_S	43.32 GiB	46,514,882,496	1.627	—	mradermacher
I1-IQ1_M	48.07 GiB	51,614,648,256	1.806	—	mradermacher
UD-TQ1_0	51.94 GiB	55,765,943,424	1.951	—	unsloth
I1-IQ2_XXS	55.99 GiB	60,114,257,856	2.103	—	mradermacher
UD-IQ1_S3 shards	58.86 GiB	63,202,444,768	2.211	—	unsloth
I1-IQ2_XS	62.35 GiB	66,950,514,624	2.342	—	mradermacher
I1-IQ2_S	63.36 GiB	68,036,279,232	2.380	—	mradermacher
UD-IQ1_M3 shards	63.74 GiB	68,439,344,608	2.394	—	unsloth
UD-IQ2_XXS3 shards	69.03 GiB	74,123,085,280	2.593	—	unsloth
I1-IQ2_M	69.70 GiB	74,835,966,912	2.618	—	mradermacher
I1-Q2_K_S	72.23 GiB	77,559,270,336	2.713	—	mradermacher
UD-IQ2_M3 shards	72.83 GiB	78,199,948,736	2.735	—	unsloth
I1-Q2_K	77.58 GiB	83,302,681,536	2.914	—	mradermacher
Q2_K3 shards	77.58 GiB	83,302,682,080	2.914	—	unsloth
Q2_K_L3 shards	77.72 GiB	83,446,728,128	2.919	—	unsloth
I1-IQ3_XXS	81.91 GiB	87,953,652,672	3.077	—	mradermacher
UD-IQ3_XXS3 shards	86.91 GiB	93,315,761,632	3.264	—	unsloth
I1-IQ3_XS	86.92 GiB	93,329,932,224	3.265	—	mradermacher
I1-Q3_K_S	91.91 GiB	98,690,596,800	3.452	—	mradermacher
Q3_K_S3 shards	91.91 GiB	98,690,597,344	3.452	—	unsloth
I1-IQ3_S	91.94 GiB	98,716,499,904	3.453	—	mradermacher
I1-IQ3_M	93.13 GiB	99,994,943,424	3.498	—	mradermacher
I1-Q3_K_M	101.77 GiB	109,272,481,728	3.822	—	mradermacher
Q3_K_M4 shards	101.77 GiB	109,272,482,368	3.822	—	unsloth
I1-Q3_K_L	110.22 GiB	118,351,052,736	4.140	—	mradermacher
I1-IQ4_XS	113.51 GiB	121,879,917,504	4.263	—	mradermacher
IQ4_XS4 shards	113.53 GiB	121,899,124,320	4.264	—	unsloth
IQ4_NL4 shards	120.14 GiB	128,999,622,208	4.512	—	unsloth
I1-Q4_0	120.61 GiB	129,503,724,480	4.530	—	mradermacher
Q4_04 shards	120.61 GiB	129,503,725,120	4.530	—	unsloth
I1-Q4_K_S	121.10 GiB	130,033,779,648	4.549	—	mradermacher
Q4_K_S4 shards	121.10 GiB	130,033,780,320	4.549	—	unsloth
Q4_K_M4 shards	128.84 GiB	138,342,384,064	4.839	—	lmstudio-community
I1-Q4_K_M	128.84 GiB	138,342,384,576	4.839	—	mradermacher
Q4_K_M4 shards	128.84 GiB	138,342,385,216	4.839	—	unsloth
I1-Q4_1	133.39 GiB	143,226,840,000	5.010	—	mradermacher
Q4_14 shards	133.39 GiB	143,226,840,672	5.010	—	unsloth
I1-Q5_K_S	146.66 GiB	157,478,437,824	5.509	—	mradermacher
Q5_K_S5 shards	146.66 GiB	157,478,438,592	5.509	—	unsloth
I1-Q5_K_M	151.16 GiB	162,303,935,424	5.677	—	mradermacher

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.97 GiB	0.97 GiB	—	62 / 0 / 0
8,192	1.94 GiB	1.94 GiB	—	62 / 0 / 0
16,384	3.88 GiB	3.88 GiB	—	62 / 0 / 0
32,768	7.75 GiB	7.75 GiB	—	62 / 0 / 0
65,536	15.50 GiB	15.50 GiB	—	62 / 0 / 0
131,072	31.00 GiB	31.00 GiB	—	62 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs DeepSeek-V4-Flash vs Qwen3-30B-A3B-Thinking-2507 vs Laguna-XS-2.1 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 119.81 GiB. The real file is 128.84 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

3072

Vocab

200,064

Sliding window

none

SWA period

—

MLA

Experts

256

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does MiniMax-M2.5 need?: Q4_K_M is exactly 138,342,384,064 bytes (128.84 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is MiniMax-M2.5's KV cache?: 7.75 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Is MiniMax-M2.5 a mixture-of-experts model?: Yes — 256 experts, 8 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of MiniMax-M2.5 should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.