MiniMaxAI · vision language · mixture of experts

MiniMax-M3

MiniMaxAI/MiniMax-M3

MiniMax-M3 at Q4_K_M is exactly 261,282,110,272 bytes (243.34 GiB / 261.28 GB) — an effective 4.895 bits per weight, not the nominal 4. Its KV cache at 32K is 3.75 GiB.

From the file· summed from 7 file(s)From the file· KV per layer
Parameters
427B
total, not active
Architecture
minimax-m3
60 layers
Context
1,048,576
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
IQ1_S3 shards84.31 GiB90,527,490,4001.696—bartowski
IQ1_M3 shards93.82 GiB100,738,523,4881.887—bartowski
IQ2_XXS3 shards108.60 GiB116,611,867,0082.185—bartowski
UD-IQ1_M4 shards119.60 GiB128,417,281,1522.406—unsloth
IQ2_XS4 shards120.63 GiB129,521,934,8162.426—bartowski
IQ2_S4 shards123.06 GiB132,138,191,3282.475—bartowski
UD-IQ2_XXS4 shards124.80 GiB134,004,094,0802.510—unsloth
UD-IQ2_M4 shards124.99 GiB134,208,566,4002.514—unsloth
IQ2_M4 shards135.58 GiB145,576,741,3442.727—bartowski
Q2_K4 shards141.46 GiB151,886,604,7362.845—bartowski
Q2_K_L4 shards142.57 GiB153,086,988,7682.868—bartowski
UD-IQ3_XXS5 shards148.46 GiB159,407,162,6242.986—unsloth
UD-IQ3_S5 shards162.75 GiB174,752,023,8083.274—unsloth
IQ3_XXS5 shards167.65 GiB180,011,108,9603.372—bartowski
Q3_K_S5 shards174.39 GiB187,251,069,5043.508—bartowski
UD-Q3_K_M5 shards181.22 GiB194,578,416,8963.645—unsloth
IQ3_XS5 shards182.87 GiB196,357,952,0963.679—bartowski
Q3_K_M5 shards183.11 GiB196,614,525,5363.683—bartowski
Q3_K_L6 shards190.89 GiB204,971,152,1283.840—bartowski
IQ3_M6 shards191.41 GiB205,523,227,3923.850—bartowski
UD-IQ4_XS6 shards193.31 GiB207,559,687,5203.888—unsloth
UD-IQ4_NL6 shards197.24 GiB211,787,545,9843.967—unsloth
IQ4_XS6 shards213.91 GiB229,685,754,5924.303—bartowski
IQ4_NL7 shards226.08 GiB242,752,199,4564.548—bartowski
Q4_07 shards226.91 GiB243,644,013,3444.564—bartowski
UD-Q4_K_S7 shards230.71 GiB247,724,342,7844.641—unsloth
Q4_K_S7 shards234.10 GiB251,358,911,3284.709—bartowski
Q4_K_M7 shards243.34 GiB261,282,110,2724.895—bartowski
UD-Q4_K_M7 shards245.90 GiB264,031,796,7364.946—unsloth
Q4_17 shards250.43 GiB268,894,526,3045.037—bartowski
Q5_K_S8 shards274.95 GiB295,228,545,9845.531—bartowski
UD-Q5_K_S8 shards278.24 GiB298,756,339,2645.597—unsloth
Q5_K_M8 shards284.38 GiB305,348,746,1765.720—bartowski
UD-Q5_K_M8 shards296.01 GiB317,838,325,3445.954—unsloth
UD-Q6_K9 shards330.08 GiB354,416,850,6246.639—unsloth
Q6_K10 shards344.03 GiB369,400,554,6886.920—bartowski
Q8_011 shards421.62 GiB452,714,559,3928.481—unsloth
Q8_012 shards422.46 GiB453,611,164,7688.498—bartowski
BF1618 shards793.49 GiB852,000,340,54415.961—unsloth

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.47 GiB0.47 GiB—60 / 0 / 0
8,1920.94 GiB0.94 GiB—60 / 0 / 0
16,3841.88 GiB1.88 GiB—60 / 0 / 0
32,7683.75 GiB3.75 GiB—60 / 0 / 0
65,5367.50 GiB7.50 GiB—60 / 0 / 0
131,07215.00 GiB15.00 GiB—60 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3.6-35B-A3Bvs Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPvs gemma-4-26B-A4B-itvs Qwythos-9B-Claude-Mythos-5-1Mvs gemma-4-31B-itvs Muse-Glimmer-30BAll alternatives →Will it fit on my card? →

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 223.71 GiB. The real file is 243.34 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
60
Attention heads
64
KV heads
4
Head dim
128
Hidden size
6144
Vocab
200,064
Sliding window
none
SWA period
—
MLA
no
Experts
128
Experts per token
4
use_sliding_window
—

Questions people ask

How much VRAM does MiniMax-M3 need?
Q4_K_M is exactly 261,282,110,272 bytes (243.34 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-M3's KV cache?
3.75 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-M3 a mixture-of-experts model?
Yes — 128 experts, 4 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-M3 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.