nvidia · text · mixture of experts

NVIDIA-Nemotron-3-Super-120B-A12B-BF16

nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16

NVIDIA-Nemotron-3-Super-120B-A12B-BF16 at Q4_K_M is exactly 86,051,079,584 bytes (80.14 GiB / 86.05 GB) — an effective 5.569 bits per weight, not the nominal 4. Its KV cache at 32K is 2.75 GiB.

From the file· summed from 3 file(s)From the file· KV per layer
Parameters
124B
total, not active
Architecture
nemotron_h_moe
88 layers
Context
262,144
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
IQ1_S43.19 GiB46,379,649,4403.002—bartowski
IQ1_M44.51 GiB47,795,718,5603.093—bartowski
IQ2_XXS2 shards46.71 GiB50,155,833,9843.246—bartowski
IQ2_XS2 shards48.47 GiB52,043,926,1123.368—bartowski
IQ2_S2 shards48.54 GiB52,115,229,3123.373—bartowski
UD-IQ1_M3 shards49.06 GiB52,676,904,8003.409—unsloth
UD-IQ2_XXS3 shards49.06 GiB52,676,904,8003.409—unsloth
UD-IQ2_M3 shards49.12 GiB52,744,013,6643.413—unsloth
IQ2_M2 shards50.29 GiB54,003,321,4723.495—bartowski
Q2_K2 shards51.05 GiB54,819,768,9603.548—bartowski
Q2_K_L2 shards51.54 GiB55,344,056,9283.582—bartowski
UD-IQ3_S3 shards52.74 GiB56,631,805,7923.665—unsloth
UD-IQ3_XXS3 shards52.74 GiB56,631,805,7923.665—unsloth
Q3_K_S2 shards56.93 GiB61,127,510,6563.956—bartowski
IQ3_XXS2 shards57.40 GiB61,634,038,3683.989—bartowski
UD-Q3_K_M3 shards57.47 GiB61,712,254,8163.994—unsloth
UD-Q3_K_S3 shards57.47 GiB61,712,254,8163.994—unsloth
UD-IQ4_NL3 shards60.06 GiB64,484,689,7604.173—unsloth
UD-IQ4_XS3 shards60.06 GiB64,484,689,7604.173—unsloth
IQ3_XS2 shards60.11 GiB64,540,625,5364.177—bartowski
Q3_K_M2 shards60.21 GiB64,650,726,0164.184—bartowski
IQ3_M2 shards61.76 GiB66,319,010,4324.292—bartowski
Q3_K_L2 shards61.83 GiB66,394,507,9044.297—bartowski
IQ4_XS2 shards62.59 GiB67,204,303,4564.349—bartowski
IQ4_NL2 shards64.39 GiB69,138,795,1364.475—bartowski
Q4_02 shards66.12 GiB70,993,726,0804.595—bartowski
Q4_12 shards71.37 GiB76,632,705,6644.960—bartowski
Q4_K_S2 shards72.68 GiB78,040,156,8005.051—bartowski
UD-Q4_K_S3 shards73.59 GiB79,017,953,1205.114—unsloth
UD-Q4_K_M3 shards76.87 GiB82,541,168,4805.342—unsloth
Q4_K_M3 shards80.14 GiB86,051,079,5845.569—lmstudio-community
Q4_K_M3 shards81.01 GiB86,981,954,3045.629—bartowski
Q4_K_L3 shards81.38 GiB87,380,413,1525.655—bartowski
Q5_K_S3 shards81.55 GiB87,563,389,6645.667—bartowski
UD-Q5_K_S3 shards83.56 GiB89,717,622,6245.806—unsloth
Q5_K_M3 shards89.97 GiB96,602,573,5366.252—bartowski
Q5_K_L3 shards90.28 GiB96,933,923,5526.274—bartowski
UD-Q5_K_M4 shards99.96 GiB107,333,699,5526.947—unsloth
Q6_K3 shards105.17 GiB112,921,986,4647.308—lmstudio-community
Q6_K3 shards105.76 GiB113,563,978,4967.350—bartowski

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.34 GiB0.34 GiB—88 / 0 / 0
8,1920.69 GiB0.69 GiB—88 / 0 / 0
16,3841.38 GiB1.38 GiB—88 / 0 / 0
32,7682.75 GiB2.75 GiB—88 / 0 / 0
65,5365.50 GiB5.50 GiB—88 / 0 / 0
131,07211.00 GiB11.00 GiB—88 / 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 64.76 GiB. The real file is 80.14 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
88
Attention heads
32
KV heads
2
Head dim
128
Hidden size
4096
Vocab
131,072
Sliding window
none
SWA period
—
MLA
no
Experts
512
Experts per token
22
use_sliding_window
—

Questions people ask

How much VRAM does NVIDIA-Nemotron-3-Super-120B-A12B-BF16 need?
Q4_K_M is exactly 86,051,079,584 bytes (80.14 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is NVIDIA-Nemotron-3-Super-120B-A12B-BF16's KV cache?
2.75 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 NVIDIA-Nemotron-3-Super-120B-A12B-BF16 a mixture-of-experts model?
Yes — 512 experts, 22 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 NVIDIA-Nemotron-3-Super-120B-A12B-BF16 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.