Llama 3.2 3B Instruct memory requirements

3.2B parameters — 6.0 GiB of weights at FP16, 1.9 GiB at Q4_K_M, and 112.0 KiB of KV cache per token on top. Solved against 22 real accelerators; 22 of them can hold it.

Open the calculator →All models

Parameters

3.2B

3,212,749,824 exactly

Weights, Q4_K_M

1.9 GiB

6.0 GiB unquantized

KV cache

112.0 KiB

per token · 28L × 8 KV × 128

Smallest card (Q4)

8 GB

Jetson Orin Nano Super (8GB)

Weights at each precision

Precision	Bits/weight	Weights	Source	Notes
FP16 / BF16	16.00	6.0 GiB	measured file	Unquantized. Two bytes per parameter, exactly.
Q8_0	8.52	3.2 GiB	measured file	Effectively lossless; the usual reference point for quantized quality.
Q6_K	6.58	2.5 GiB	measured file	Quality loss is hard to measure on most benchmarks.
Q5_K_M	5.78	2.2 GiB	measured file	A middle point when Q4 fits with too little room for context.
Q4_K_M	5.03	1.9 GiB	measured file	The default choice for local inference — the best size/quality knee.
Q3_K_M	3.91	1.5 GiB	nominal	Measurable quality loss. Worth it only to make a model fit at all.

Every accelerator

Largest quant that fits with at least 4k of context, the context it leaves, and the bandwidth ceiling on decode speed. Each row links to the worked page for that pairing.

Accelerator	Memory	Bandwidth	Best quant	Max context	Ceiling
Apple M3 Ultra (512GB)	512 GB	819 GB/s	FP16 / BF16	131,072	111 tok/s
H200 141GB (SXM)	141 GB	4800 GB/s	FP16 / BF16	131,072	652 tok/s
Apple M4 Max (128GB)	128 GB	546 GB/s	FP16 / BF16	131,072	74 tok/s
H100 80GB (SXM5)	80 GB	3350 GB/s	FP16 / BF16	131,072	455 tok/s
A100 80GB (SXM)	80 GB	2039 GB/s	FP16 / BF16	131,072	277 tok/s
Jetson AGX Orin (64GB)	64 GB	204.8 GB/s	FP16 / BF16	131,072	28 tok/s
RTX 6000 Ada Generation	48 GB	960 GB/s	FP16 / BF16	131,072	130 tok/s
L40S	48 GB	864 GB/s	FP16 / BF16	131,072	117 tok/s
Apple M4 Pro (48GB)	48 GB	273 GB/s	FP16 / BF16	131,072	37 tok/s
A100 40GB (SXM)	40 GB	1555 GB/s	FP16 / BF16	131,072	211 tok/s
GeForce RTX 5090	32 GB	1792 GB/s	FP16 / BF16	131,072	243 tok/s
GeForce RTX 4090	24 GB	1008 GB/s	FP16 / BF16	131,072	137 tok/s
Radeon RX 7900 XTX	24 GB	960 GB/s	FP16 / BF16	131,072	130 tok/s
GeForce RTX 3090	24 GB	936 GB/s	FP16 / BF16	131,072	127 tok/s
Apple M4 (24GB)	24 GB	120 GB/s	FP16 / BF16	112,494	16 tok/s
GeForce RTX 5080	16 GB	960 GB/s	FP16 / BF16	81,786	130 tok/s
GeForce RTX 5070 Ti	16 GB	896 GB/s	FP16 / BF16	81,786	122 tok/s
GeForce RTX 4080 SUPER	16 GB	736 GB/s	FP16 / BF16	81,786	100 tok/s
GeForce RTX 4070 Ti SUPER	16 GB	672 GB/s	FP16 / BF16	81,786	91 tok/s
GeForce RTX 4060 Ti 16GB	16 GB	288 GB/s	FP16 / BF16	81,786	39 tok/s
GeForce RTX 3060 12GB	12 GB	360 GB/s	FP16 / BF16	47,333	49 tok/s
Jetson Orin Nano Super (8GB)	8 GB	102 GB/s	Q8_0	26,337	23 tok/s

Questions about this model

How much VRAM does Llama 3.2 3B Instruct need?

6.0 GiB for the weights at FP16 and 1.9 GiB at Q4_K_M, from an exact count of 3,212,749,824 parameters. On top of that, the KV cache costs 112.0 KiB per token of context — 896 MiB at 8k and 3.5 GiB at 32k.

What is the smallest GPU that runs Llama 3.2 3B Instruct?

Of the 22 accelerators here, the smallest that holds Q4_K_M weights is the Jetson Orin Nano Super (8GB) at 8 GB — 1.9 GiB of weights against 6.0 GiB usable, leaving room for 38k tokens. For unquantized weights you need at least a Jetson Orin Nano Super (8GB).

Why is the KV cache for Llama 3.2 3B Instruct the size it is?

Because it stores one key and one value vector per token, per layer: 28 layers × 8 KV heads × 128 dimensions × 2 (K and V) × 2 bytes = 112.0 KiB per token. Grouped-query attention shares those 8 KV heads across 24 query heads, cutting the cache by 3× against multi-head attention. None of this can be read off the parameter count.

Compare with

Method and limits. The parameter count and every architecture figure on this page are read from the model's own published config and the Hub's index over its tensor shapes, fetched 2026-08-07 — nothing here is recalled from a model's name. Weight bytes are the size of a real published file wherever one exists and the parameter count times the published llama.cpp bits-per-weight where it does not; every row says which. Speed figures are roofline bounds — memory bandwidth divided by bytes read per token — not benchmarks. Nothing on this page is written by a language model.