Llama on NVIDIA Blackwell

Can I run Llama 3.2 3B on an NVIDIA DGX Spark (GB10)?

Yes. Llama 3.2 3B at Q4_K_M uses 3.2 GB of the 96.0 GB available on a NVIDIA DGX Spark (GB10), and runs at about 78.1 tokens per second. You can push the context to 128K.

Runs comfortably

3.2 GB of 96.0 GB · 3%
096 GB
Weights 1.8 GB
KV cache 0.9 GB
Runtime overhead 0.5 GB

Llama 3.2 3B at Q4_K_M leaves 92.8 GB spare on a NVIDIA DGX Spark (GB10). There is room to raise the context length or move up a quantisation level.

Generation78.1tok/s
Prompt processing8178tok/s
Max context128Ktokens
KV per 1K tokens0GB

Every quantisation of Llama 3.2 3B on a NVIDIA DGX Spark (GB10)

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 6.0 GB 7.4 GB Runs comfortably 128K 28.0
INT8 / W8A8 8.50 3.1 GB 4.5 GB Runs comfortably 128K 50.8
Q8_0 (GGUF) 8.50 3.1 GB 4.5 GB Runs comfortably 128K 50.8
FP8 (E4M3) 8.00 3.0 GB 4.4 GB Runs comfortably 128K 52.2
Q6_K 6.56 2.5 GB 3.8 GB Runs comfortably 128K 61.8
AWQ 4-bit 4.25 2.1 GB 3.5 GB Runs comfortably 128K 69.5
GPTQ 4-bit 4.25 2.1 GB 3.5 GB Runs comfortably 128K 69.5
MXFP4 4.25 2.1 GB 3.5 GB Runs comfortably 128K 69.5
Q5_K_M 5.67 2.1 GB 3.5 GB Runs comfortably 128K 69.8
Q5_K_S 5.52 2.1 GB 3.4 GB Runs comfortably 128K 71.3
Q4_K_M 4.85 1.8 GB 3.2 GB Runs comfortably 128K 78.1
Q4_K_S 4.58 1.8 GB 3.1 GB Runs comfortably 128K 81.2
Q4_0 4.55 1.7 GB 3.1 GB Runs comfortably 128K 81.6
IQ4_XS 4.25 1.6 GB 3.0 GB Runs comfortably 128K 85.4
Q3_K_M 3.91 1.5 GB 2.9 GB Runs comfortably 128K 90.1
IQ3_M 3.70 1.5 GB 2.8 GB Runs comfortably 128K 93.3
IQ3_XXS 3.06 1.3 GB 2.6 GB Runs comfortably 128K 105
Q2_K 2.63 1.1 GB 2.5 GB Runs comfortably 128K 114
IQ2_XXS 2.06 0.9 GB 2.3 GB Runs comfortably 128K 129
IQ1_M 1.75 0.8 GB 2.2 GB Runs comfortably 128K 139

Also worth checking

Llama 3.2 3B on every GPU
Full hardware tableEverything that runs on a NVIDIA DGX Spark (GB10)
Full model tableWhy the KV cache is the part that bites
Guide