Llama · 1.2B parameters

Llama 3.2 1B VRAM requirements

Llama 3.2 1B has 16 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 0.7 GB, and the best quantisation that fits a 24 GB card is FP16 / BF16.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Runs comfortably

1.8 GB of 21.8 GB · 8%

022 GB

Weights 0.7 GB

KV cache 0.3 GB

Runtime overhead 0.8 GB

Llama 3.2 1B at Q4_K_M leaves 20.0 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation669tok/s

Prompt processing27944tok/s

Max context128Ktokens

KV per 1K tokens0GB

Every quantisation of Llama 3.2 1B on a RTX 4090

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

Quantisation	bpw	Weights	Total	Verdict	Max ctx	tok/s
FP16 / BF16	16.00	2.3 GB	3.3 GB	Runs comfortably	128K	242
INT8 / W8A8	8.50	1.2 GB	2.2 GB	Runs comfortably	128K	448
Q8_0 (GGUF)	8.50	1.2 GB	2.2 GB	Runs comfortably	128K	448
FP8 (E4M3)	8.00	1.2 GB	2.2 GB	Runs comfortably	128K	452
AWQ 4-bit	4.25	1.0 GB	2.0 GB	Runs comfortably	128K	523
GPTQ 4-bit	4.25	1.0 GB	2.0 GB	Runs comfortably	128K	523
MXFP4	4.25	1.0 GB	2.0 GB	Runs comfortably	128K	523
Q6_K	6.56	0.9 GB	2.0 GB	Runs comfortably	128K	535
Q5_K_M	5.67	0.8 GB	1.9 GB	Runs comfortably	128K	604
Q5_K_S	5.52	0.8 GB	1.8 GB	Runs comfortably	128K	617
Q4_K_M	4.85	0.7 GB	1.8 GB	Runs comfortably	128K	669
Q4_K_S	4.58	0.7 GB	1.7 GB	Runs comfortably	128K	693
Q4_0	4.55	0.7 GB	1.7 GB	Runs comfortably	128K	696
IQ4_XS	4.25	0.7 GB	1.7 GB	Runs comfortably	128K	724
Q3_K_M	3.91	0.6 GB	1.6 GB	Runs comfortably	128K	759
IQ3_M	3.70	0.6 GB	1.6 GB	Runs comfortably	128K	783
IQ3_XXS	3.06	0.5 GB	1.6 GB	Runs comfortably	128K	864
Q2_K	2.63	0.5 GB	1.5 GB	Runs comfortably	128K	929
IQ2_XXS	2.06	0.4 GB	1.4 GB	Runs comfortably	128K	1032
IQ1_M	1.75	0.4 GB	1.4 GB	Runs comfortably	128K	1098

Llama 3.2 1B on each GPU

Q4_K_M weights at 8K context, single card, monitor attached.

GPU	VRAM	GB/s	Verdict	Max ctx	tok/s
H100 SXM 80GB	80	3350	Runs comfortably	128K	2192
A100 80GB	80	2039	Runs comfortably	128K	1319
RTX 5090	32	1792	Runs comfortably	128K	1235
RTX 5080	16	960	Runs comfortably	128K	700
RTX 4090	24	1008	Runs comfortably	128K	669
RTX 5070 Ti	16	896	Runs comfortably	128K	657
RTX 3090	24	936	Runs comfortably	128K	650
Radeon RX 7900 XTX	24	960	Runs comfortably	128K	609
L40S	48	864	Runs comfortably	128K	579
RTX A6000	48	768	Runs comfortably	128K	539
RTX 3080 10GB	10	760	Runs comfortably	128K	534
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	531
RTX 5070	12	672	Runs comfortably	128K	500
RTX 4080 Super	16	736	Runs comfortably	128K	497
RTX 4070 Ti Super	16	672	Runs comfortably	128K	456
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	400
RTX 4070 Super	12	504	Runs comfortably	128K	346
RTX 4070	12	504	Runs comfortably	128K	346
RTX 5060 Ti 16GB	16	448	Runs comfortably	128K	339
Arc B580	12	456	Runs comfortably	128K	267
RTX 3060 12GB	12	360	Runs comfortably	128K	260
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	209
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	128K	204
RTX 4060 Ti 16GB	16	288	Runs comfortably	128K	200
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	157

Architecture

Parameters	1.2B
Layers	16
Hidden size	2048
Attention heads / KV heads	32 / 8
Head dimension	64
Vocabulary	128,256
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	meta-llama/Llama-3.2-1B-Instruct

Direct answers

Llama 3.2 1B on RTX 5090
See the verdict Llama 3.2 1B on RTX 4090
See the verdict Llama 3.2 1B on RTX 3090
See the verdict Llama 3.2 1B on RTX 5080
See the verdict Llama 3.2 1B on RTX 5070 Ti
See the verdict Llama 3.2 1B on RTX 5070
See the verdict Llama 3.2 1B on RTX 4070 Ti Super
See the verdict Llama 3.2 1B on RTX 4070 Super
See the verdict Llama 3.2 1B on RTX 4070
See the verdict