llmfan46 · text · mixture of experts

MiniMax-M2.7-BF16-ultra-uncensored-heretic

llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic

MiniMax-M2.7-BF16-ultra-uncensored-heretic at Q4_K_M is exactly 138,342,384,352 bytes (128.84 GiB / 138.34 GB) — an effective 4.840 bits per weight, not the nominal 4. Its KV cache at 32K is 7.75 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
229B
total, not active
Architecture
minimax-m2
62 layers
Context
204,800
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
I1-IQ1_S43.32 GiB46,514,883,0721.627—mradermacher
I1-IQ1_M48.07 GiB51,614,648,8321.806—mradermacher
I1-IQ2_XXS55.99 GiB60,114,258,4322.103—mradermacher
I1-IQ2_XS62.35 GiB66,950,515,2002.342—mradermacher
I1-IQ2_S63.36 GiB68,036,279,8082.380—mradermacher
I1-IQ2_M69.70 GiB74,835,967,4882.618—mradermacher
I1-Q2_K_S72.23 GiB77,559,270,9122.713—mradermacher
I1-Q2_K77.58 GiB83,302,682,1122.914—mradermacher
I1-IQ3_XXS81.91 GiB87,953,653,2483.077—mradermacher
I1-IQ3_XS86.92 GiB93,329,932,8003.265—mradermacher
Q3_K_S91.91 GiB98,690,596,5763.452—llmfan46
I1-Q3_K_S91.91 GiB98,690,597,3763.452—mradermacher
I1-IQ3_S91.94 GiB98,716,500,4803.453—mradermacher
Q3_K_M101.77 GiB109,272,481,5043.823—llmfan46
Q3_K_L110.22 GiB118,351,052,5124.140—llmfan46
Q4_K_S121.10 GiB130,033,779,4244.549—llmfan46
Q4_K_M128.84 GiB138,342,384,3524.840—llmfan46
Q5_K_S146.66 GiB157,478,437,6005.509—llmfan46
Q5_K_M151.16 GiB162,303,935,2005.678—llmfan46
Q6_K174.87 GiB187,763,082,9766.568—llmfan46

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.97 GiB0.97 GiB—62 / 0 / 0
8,1921.94 GiB1.94 GiB—62 / 0 / 0
16,3843.88 GiB3.88 GiB—62 / 0 / 0
32,7687.75 GiB7.75 GiB—62 / 0 / 0
65,53615.50 GiB15.50 GiB—62 / 0 / 0
131,07231.00 GiB31.00 GiB—62 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs DeepSeek-V4-Flashvs Qwen3-30B-A3B-Thinking-2507vs Laguna-XS-2.1All 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 119.80 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
62
Attention heads
48
KV heads
8
Head dim
128
Hidden size
3072
Vocab
200,064
Sliding window
none
SWA period
—
MLA
no
Experts
256
Experts per token
8
use_sliding_window
—

Questions people ask

How much VRAM does MiniMax-M2.7-BF16-ultra-uncensored-heretic need?
Q4_K_M is exactly 138,342,384,352 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-M2.7-BF16-ultra-uncensored-heretic'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-M2.7-BF16-ultra-uncensored-heretic 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-M2.7-BF16-ultra-uncensored-heretic 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.