nvidia · text

Llama-3_3-Nemotron-Super-49B-v1

nvidia/Llama-3_3-Nemotron-Super-49B-v1

Llama-3_3-Nemotron-Super-49B-v1 at Q4_K_M is exactly 30,215,576,224 bytes (28.14 GiB / 30.22 GB) — an effective 4.847 bits per weight, not the nominal 4. Its KV cache at 32K is 80.00 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
49.9B
Architecture
deci
80 layers
Context
131,072
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
IQ1_S10.27 GiB11,029,863,3281.770—mmnga
UD-IQ1_S10.66 GiB11,446,492,6721.836—unsloth
IQ1_M11.19 GiB12,015,852,4481.928—mmnga
UD-IQ1_M11.52 GiB12,373,302,7841.985—unsloth
IQ2_XXS12.72 GiB13,659,167,6482.191—mmnga
IQ2_XXS12.72 GiB13,659,167,6802.191—bartowski
UD-IQ2_XXS12.99 GiB13,950,295,5522.238—unsloth
IQ2_XS14.04 GiB15,076,580,2562.419—mmnga
IQ2_XS14.04 GiB15,076,580,2882.419—bartowski
IQ2_S14.76 GiB15,848,479,6482.542—mmnga
IQ2_S14.76 GiB15,848,479,6802.542—bartowski
IQ2_M15.98 GiB17,163,131,8082.753—mmnga
IQ2_M15.98 GiB17,163,131,8402.753—bartowski
UD-IQ2_M16.15 GiB17,341,914,6242.782—unsloth
Q2_K17.45 GiB18,739,796,6403.006—mmnga
Q2_K17.45 GiB18,739,796,9283.006—bartowski
Q2_K17.45 GiB18,739,797,5043.006—unsloth
Q2_K_L17.68 GiB18,986,049,0243.046—unsloth
IQ3_XXS18.18 GiB19,519,019,9363.131—mmnga
IQ3_XXS18.18 GiB19,519,019,9683.131—bartowski
UD-IQ3_XXS18.34 GiB19,692,428,8003.159—unsloth
Q2_K_L18.41 GiB19,765,844,9283.171—bartowski
IQ3_XS19.47 GiB20,908,006,3043.354—mmnga
IQ3_XS19.47 GiB20,908,006,3363.354—bartowski
Q3_K_S20.45 GiB21,955,336,8643.522—mmnga
IQ3_S20.45 GiB21,955,337,1203.522—mmnga
Q3_K_S20.45 GiB21,955,337,1523.522—bartowski
Q3_K_S20.45 GiB21,955,337,7283.522—unsloth
IQ3_M21.10 GiB22,657,227,6803.635—mmnga
IQ3_M21.10 GiB22,657,227,7123.635—bartowski
Q3_K_M22.64 GiB24,311,224,9923.900—mmnga
Q3_K_M22.64 GiB24,311,225,2803.900—bartowski
Q3_K_M22.64 GiB24,311,225,8563.900—unsloth
Q3_K_L24.47 GiB26,272,062,1124.215—mmnga
Q3_K_L24.47 GiB26,272,062,4004.215—bartowski
IQ4_XS25.03 GiB26,871,405,4724.311—mmnga
IQ4_XS25.03 GiB26,871,405,5044.311—bartowski
IQ4_XS25.06 GiB26,904,239,6164.316—unsloth
Q4_026.39 GiB28,332,661,4084.545—mmnga
IQ4_NL26.43 GiB28,384,041,8884.553—mmnga

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,09610.00 GiB10.00 GiB—80 / 0 / 0
8,19220.00 GiB20.00 GiB—80 / 0 / 0
16,38440.00 GiB40.00 GiB—80 / 0 / 0
32,76880.00 GiB80.00 GiB—80 / 0 / 0
65,536160.00 GiB160.00 GiB—80 / 0 / 0
131,072320.00 GiB320.00 GiB—80 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs gemma-4-12B-it-qat-q4_0-unquantizedvs gemma-4-E4B-itvs Qwen3-30B-A3B-Thinking-2507All 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 26.12 GiB. The real file is 28.14 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
80
Attention heads
64
KV heads
64
Head dim
128
Hidden size
8192
Vocab
128,256
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does Llama-3_3-Nemotron-Super-49B-v1 need?
Q4_K_M is exactly 30,215,576,224 bytes (28.14 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Llama-3_3-Nemotron-Super-49B-v1's KV cache?
80.00 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.
Which quantization of Llama-3_3-Nemotron-Super-49B-v1 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.