Llama-3.2-3B-Instruct

Meta · 3.2B · ctx 128K · llama3.2 · HuggingFace

Llama-3.2-3B-Instruct (Meta) is a 3.2B parameters open-weight model tracked on AI Hashrate. At 8K context we estimate roughly 3.4 GB VRAM for Q4 (weights 1.76 GB + KV 0.64 GB + 1 GB overhead) and 8.04 GB for FP16. About 50 GPUs/accelerators in our catalog fully fit this model at Q4 under that context assumption. Fastest Q4 configs: B200 192GB SXM ≈ 1590.9 tok/s (estimated); MI300X 192GB ≈ 1054.0 tok/s (estimated); H200 141GB SXM5 ≈ 954.5 tok/s (estimated); H100 80GB SXM5 ≈ 666.2 tok/s (estimated); Gaudi 3 128GB ≈ 651.5 tok/s (estimated). Use the table below for VRAM fit and tok/s per dollar (list MSRP). Estimates follow memory-bandwidth math; measured rows override when present. See methodology for details. Methodology.

100 hardware configs — sorted by tok/s. Default fit context 8K.

Hardware	VRAM	Quant	tok/s	Fits?	tok/s/$
B200 192GB SXM	192	Q4	1590.9 est.	Yes	0.04
MI300X 192GB	192	Q4	1054.0 est.	Yes	0.07
H200 141GB SXM5	141	Q4	954.5 est.	Yes	0.027
H100 80GB SXM5	80	Q4	666.2 est.	Yes	0.022
Gaudi 3 128GB	128	Q4	651.5 est.	Yes	0.052
B200 192GB SXM	192	FP16	437.5 est.	Yes	0.011
A100 80GB SXM4	80	Q4	405.5 est.	Yes	0.027
RTX 5090 32GB	32	Q4	356.4 est.	Yes	0.178
MI300X 192GB	192	FP16	289.8 est.	Yes	0.019
H200 141GB SXM5	141	FP16	262.5 est.	Yes	0.007
RTX 4090 24GB	24	Q4	200.5 est.	Yes	0.125
RTX 4090D 24GB (China)	24	Q4	200.5 est.	Yes	0.125
RTX 3090 Ti 24GB	24	Q4	200.5 est.	Yes	0.1
RTX 4070 Ti Super 16GB	16	Q4	199.0 measured	Yes	0.249
RTX 5080 16GB	16	Q4	190.9 est.	Yes	0.191
RX 7900 XTX 24GB	24	Q4	190.9 est.	Yes	0.191
RTX 6000 Ada 48GB	48	Q4	190.9 est.	Yes	0.028
RTX 3090 24GB	24	Q4	186.1 est.	Yes	0.124
H100 80GB SXM5	80	FP16	183.2 est.	Yes	0.006
Gaudi 3 128GB	128	FP16	179.2 est.	Yes	0.014
RTX 5070 Ti 16GB	16	Q4	178.2 est.	Yes	0.238
L40S 48GB	48	Q4	171.8 est.	Yes	0.021
RTX 4070 12GB	12	Q4	160.7 measured	Yes	0.268
Mac Studio M2 Ultra 192GB	192	Q4	159.1 est.	Yes	0.023
Mac Studio M3 Ultra 96GB	96	Q4	154.6 measured	Yes	0.039
RTX A6000 48GB	48	Q4	152.7 est.	Yes	0.034
RTX 3080 10GB	10	Q4	151.1 est.	Yes	0.216
RTX 4080 Super 16GB	16	Q4	146.4 est.	Yes	0.147
RTX 5070 12GB	12	Q4	133.6 est.	Yes	0.243
RX 7900 XT 20GB	20	Q4	131.2 measured	Yes	0.146
RX 9070 16GB	16	Q4	127.3 est.	Yes	0.232
RX 7800 XT 16GB	16	Q4	124.1 est.	Yes	0.249
RTX 3060 12GB	12	Q4	122.9 measured	Yes	0.374
RX 7900 GRE 16GB	16	Q4	114.5 est.	Yes	0.209
A100 80GB SXM4	80	FP16	111.5 est.	Yes	0.007
Arc A770 16GB	16	Q4	111.4 est.	Yes	0.319
RTX 4070 Ti 12GB	12	Q4	110.2 est.	Yes	0.138
MacBook Pro M4 Max 128GB	128	Q4	108.6 est.	Yes	0.025
RTX 4070 Super 12GB	12	Q4	100.2 est.	Yes	0.167
RTX 5090 32GB	32	FP16	98.0 est.	Yes	0.049
RX 9070 XT 16GB	16	Q4	89.6 measured	Yes	0.15
RTX 5060 8GB	8	Q4	89.1 est.	Yes	0.298
RTX 5060 Ti 8GB	8	Q4	89.1 est.	Yes	0.235
RTX 5060 Ti 16GB	16	Q4	89.1 est.	Yes	0.208
RTX 3070 8GB	8	Q4	89.1 est.	Yes	0.179
RX 7700 XT 12GB	12	Q4	85.9 est.	Yes	0.191
MacBook Pro M3 Max 128GB	128	Q4	79.5 est.	Yes	0.019
Arc B570 10GB	10	Q4	75.6 est.	Yes	0.345
RTX 5060 8GB	8	FP16	69.3 est.	No	0.232
RTX 5060 Ti 8GB	8	FP16	69.3 est.	No	0.183
RTX 3070 8GB	8	FP16	69.3 est.	No	0.139
T4 16GB	16	Q4	63.6 est.	Yes	0.032
RTX 4060 Ti 8GB	8	Q4	57.3 est.	Yes	0.144
RTX 4060 Ti 16GB	16	Q4	57.3 est.	Yes	0.115
RX 7600 XT 16GB	16	Q4	57.3 est.	Yes	0.174
RTX 4090 24GB	24	FP16	55.1 est.	Yes	0.034
RTX 4090D 24GB (China)	24	FP16	55.1 est.	Yes	0.034
RTX 3090 Ti 24GB	24	FP16	55.1 est.	Yes	0.028
Mac mini M4 Pro 64GB	64	Q4	54.3 est.	Yes	0.025
RTX 4060 8GB	8	Q4	54.1 est.	Yes	0.181
RTX 5080 16GB	16	FP16	52.5 est.	Yes	0.053
RX 7900 XTX 24GB	24	FP16	52.5 est.	Yes	0.053
RTX 6000 Ada 48GB	48	FP16	52.5 est.	Yes	0.008
RTX 3090 24GB	24	FP16	51.2 est.	Yes	0.034
RTX 5070 Ti 16GB	16	FP16	49.0 est.	Yes	0.065
RX 7600 8GB	8	Q4	48.5 measured	Yes	0.18
Arc B580 12GB	12	Q4	47.4 measured	Yes	0.19
L40S 48GB	48	FP16	47.2 est.	Yes	0.006
Mac Studio M3 Ultra 96GB	96	FP16	44.8 est.	Yes	0.011
RTX 4060 Ti 8GB	8	FP16	44.6 est.	No	0.112
RX 7600 8GB	8	FP16	44.6 est.	No	0.166
Mac Studio M2 Ultra 192GB	192	FP16	43.8 est.	Yes	0.006
RX 7900 XT 20GB	20	FP16	43.8 est.	Yes	0.049
RTX 4060 8GB	8	FP16	42.1 est.	No	0.141
RTX A6000 48GB	48	FP16	42.0 est.	Yes	0.009
RTX 3080 10GB	10	FP16	41.6 est.	Yes	0.06
Mac mini M4 32GB	32	Q4	41.3 measured	Yes	0.041
RTX 4080 Super 16GB	16	FP16	40.2 est.	Yes	0.04
RTX 4070 Ti Super 16GB	16	FP16	36.8 est.	Yes	0.046
RTX 5070 12GB	12	FP16	36.8 est.	Yes	0.067
RX 9070 16GB	16	FP16	35.0 est.	Yes	0.064
RX 9070 XT 16GB	16	FP16	35.0 est.	Yes	0.058
RX 7800 XT 16GB	16	FP16	34.1 est.	Yes	0.068
RX 7900 GRE 16GB	16	FP16	31.5 est.	Yes	0.057
Arc A770 16GB	16	FP16	30.6 est.	Yes	0.088
RTX 4070 Ti 12GB	12	FP16	30.3 est.	Yes	0.038
MacBook Pro M4 Max 128GB	128	FP16	29.9 est.	Yes	0.007
RTX 4070 12GB	12	FP16	27.6 est.	Yes	0.046
RTX 4070 Super 12GB	12	FP16	27.6 est.	Yes	0.046
Arc B580 12GB	12	FP16	24.9 est.	Yes	0.1
RTX 5060 Ti 16GB	16	FP16	24.5 est.	Yes	0.057
RX 7700 XT 12GB	12	FP16	23.6 est.	Yes	0.053
MacBook Pro M3 Max 128GB	128	FP16	21.9 est.	Yes	0.005
Arc B570 10GB	10	FP16	20.8 est.	Yes	0.095
RTX 3060 12GB	12	FP16	19.7 est.	Yes	0.06
T4 16GB	16	FP16	17.5 est.	Yes	0.009
RTX 4060 Ti 16GB	16	FP16	15.7 est.	Yes	0.031
RX 7600 XT 16GB	16	FP16	15.7 est.	Yes	0.048
Mac mini M4 Pro 64GB	64	FP16	14.9 est.	Yes	0.007
Mac mini M4 32GB	32	FP16	6.6 est.	Yes	0.007

Fit checks

• RTX 5090 32GB — fits · ≈356.4 tok/s Q4 est.
• RTX 3090 Ti 24GB — fits · ≈200.5 tok/s Q4 est.
• RTX 4090 24GB — fits · ≈200.5 tok/s Q4 est.
• RTX 5080 16GB — fits · ≈190.9 tok/s Q4 est.
• RTX 6000 Ada 48GB — fits · ≈190.9 tok/s Q4 est.
• RX 7900 XTX 24GB — fits · ≈190.9 tok/s Q4 est.
• RTX 3090 24GB — fits · ≈186.1 tok/s Q4 est.
• RTX 5070 Ti 16GB — fits · ≈178.2 tok/s Q4 est.
• L40S 48GB — fits · ≈171.8 tok/s Q4 est.
• Mac Studio M3 Ultra 96GB — fits · ≈154.6 tok/s Q4 measured
• Mac Studio M2 Ultra 192GB — fits · ≈159.1 tok/s Q4 est.
• RX 7900 XT 20GB — fits · ≈131.2 tok/s Q4 measured

12 popular retail GPUs; every row in the table above also links its Fits? verdict to the full fit check.

Same family: Llama

Llama 3.1 Llama-3.1-8B-Instruct · Llama-3.1-70B-Instruct
Llama 3.3 Llama-3.3-70B-Instruct
Llama 4 Llama-4-Scout-17B-16E-Instruct

All Llama models →