nvidia · text

NVIDIA-Nemotron-3-Nano-4B-BF16

nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16

NVIDIA-Nemotron-3-Nano-4B-BF16 at Q4_K_M is exactly 2,837,072,896 bytes (2.64 GiB / 2.84 GB) — an effective 5.712 bits per weight, not the nominal 4. Its KV cache at 32K is 5.25 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
4.0B
Architecture
nemotron_h
42 layers
Context
262,144
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
UD-IQ2_XXS2.03 GiB2,182,496,6724.394—unsloth
IQ2_M2.13 GiB2,286,132,0324.603—bartowski
UD-IQ2_M2.14 GiB2,298,861,0884.628—unsloth
Q3_K_S2.19 GiB2,355,735,5844.743—unsloth
UD-IQ3_XXS2.23 GiB2,389,391,1364.811—unsloth
IQ3_XXS2.25 GiB2,413,265,4724.859—bartowski
Q2_K2.26 GiB2,430,124,6084.893—bartowski
IQ3_XS2.29 GiB2,457,031,4884.947—bartowski
Q3_K_S2.29 GiB2,457,031,4884.947—bartowski
Q3_K_M2.30 GiB2,464,924,8324.963263unsloth
IQ3_M2.32 GiB2,494,092,7365.021—bartowski
Q4_02.35 GiB2,528,598,1765.091—unsloth
IQ4_XS2.37 GiB2,544,779,9365.123263unsloth
IQ4_XS2.38 GiB2,558,380,7365.151263bartowski
Q3_K_M2.39 GiB2,566,220,7365.167263bartowski
IQ4_NL2.39 GiB2,567,685,2805.170—unsloth
IQ4_NL2.41 GiB2,589,565,1205.214—bartowski
Q4_02.42 GiB2,601,858,2405.238263bartowski
Q2_K_L2.45 GiB2,635,645,5045.306—bartowski
Q3_K_L2.46 GiB2,640,380,8645.316—bartowski
Q4_12.52 GiB2,707,268,6405.451—unsloth
Q4_12.60 GiB2,795,079,7445.627—bartowski
Q4_K_S2.64 GiB2,831,999,9045.702—unsloth
Q4_K_M2.64 GiB2,837,072,8965.712263lmstudio-community
Q4_K_M2.70 GiB2,900,295,7125.839263unsloth
Q4_K_S2.70 GiB2,900,844,4805.840—bartowski
Q4_K_M2.77 GiB2,978,372,6725.996263bartowski
Q5_K_S2.90 GiB3,110,558,2406.263—unsloth
Q4_K_L2.92 GiB3,132,513,3446.307—bartowski
Q5_K_S2.93 GiB3,143,740,2246.329—bartowski
Q5_K_M2.94 GiB3,159,586,4646.361263unsloth
Q5_K_M2.99 GiB3,205,312,4486.453263bartowski
Q5_K_L3.10 GiB3,333,763,0086.712—bartowski
Q6_K3.66 GiB3,933,669,3767.920—lmstudio-community
Q6_K_L3.73 GiB4,007,557,6968.069—bartowski
Q6_K3.73 GiB4,007,557,6968.069—bartowski
Q6_K3.78 GiB4,056,162,5928.166263unsloth
Q8_03.94 GiB4,233,677,9528.524—lmstudio-community
Q8_03.94 GiB4,233,679,0088.524263unsloth
Q8_03.94 GiB4,233,682,1128.524263bartowski

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.66 GiB0.66 GiB—42 / 0 / 0
8,1921.31 GiB1.31 GiB—42 / 0 / 0
16,3842.63 GiB2.63 GiB—42 / 0 / 0
32,7685.25 GiB5.25 GiB—42 / 0 / 0
65,53610.50 GiB10.50 GiB—42 / 0 / 0
131,07221.00 GiB21.00 GiB—42 / 0 / 0

Compare with

same modality, comparable size
vs gemma-4-E4B-itvs Qwen3-4Bvs Qwen3-8Bvs HyperCLOVAX-SEED-Text-Instruct-1.5Bvs Llama-3.2-1B-Instructvs Qwen3-VL-8B-Instruct-abliterated-v1All 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 2.08 GiB. The real file is 2.64 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
42
Attention heads
40
KV heads
8
Head dim
128
Hidden size
3136
Vocab
131,072
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does NVIDIA-Nemotron-3-Nano-4B-BF16 need?
Q4_K_M is exactly 2,837,072,896 bytes (2.64 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-Nano-4B-BF16's KV cache?
5.25 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 NVIDIA-Nemotron-3-Nano-4B-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.