What GPU do you need to run Llama 3.2 3B?
3.21B params 128K max context Tool use
The smallest GPU memory that runs Llama 3.2 3B well is 4 GB (Q4_K_M, e.g. T1000). 131 of 132 GPUs run it well.
MMLU 63.4
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 128K |
|---|---|---|---|---|
| Q4_K_M | 2.0 GB | 3.7 GB | 5.4 GB | 15.7 GB |
| Q8_0 | 3.4 GB | 5.1 GB | 6.8 GB | 17.1 GB |
| fp16 | 6.4 GB | 8.1 GB | 9.9 GB | 20.1 GB |
Weights plus fp16 KV cache. Leave ~1 GB headroom for the inference engine.
Cheapest GPUs that can run Llama 3.2 3B
Lowest launch MSRP among GPUs that run its smallest quantization (Q4_K_M) well, with at least 16K context.
As an Amazon Associate, I earn from qualifying purchases.
GPUs that can run Llama 3.2 3B
Best quantization that fits on each GPU, with the context and speed it gets. For chips with several memory sizes, the smallest that works is shown.
NVIDIA
- T1000 4 GB Q4_K_M 18K context · ~53 tok/s
- T400 4GB 4 GB Q4_K_M 18K context · ~26 tok/s
- T600 4 GB Q4_K_M 18K context · ~53 tok/s
- GTX 1060 6GB 6 GB Q8_0 24K context · ~43 tok/s
- GTX 1660 6 GB Q8_0 24K context · ~43 tok/s
- GTX 1660 SUPER 6 GB Q8_0 24K context · ~76 tok/s
- GTX 1660 Ti 6 GB Q8_0 24K context · ~65 tok/s
- RTX 2060 6 GB Q8_0 24K context · ~76 tok/s
- RTX 3050 6GB 6 GB Q8_0 24K context · ~38 tok/s
- RTX A2000 6 GB Q8_0 24K context · ~65 tok/s
- GTX 1070 8 GB Q8_0 42K context · ~58 tok/s
- GTX 1070 Ti 8 GB Q8_0 42K context · ~58 tok/s
- GTX 1080 8 GB Q8_0 42K context · ~72 tok/s
- RTX 2060 SUPER 8 GB Q8_0 42K context · ~101 tok/s
- RTX 2070 8 GB Q8_0 42K context · ~101 tok/s
- RTX 2070 SUPER 8 GB Q8_0 42K context · ~101 tok/s
- RTX 2080 8 GB Q8_0 42K context · ~101 tok/s
- RTX 2080 SUPER 8 GB Q8_0 42K context · ~112 tok/s
- RTX 3050 8 GB Q8_0 42K context · ~51 tok/s
- RTX 3060 Ti 8 GB Q8_0 42K context · ~101 tok/s
- RTX 3060 Ti GDDR6X 8 GB Q8_0 42K context · ~138 tok/s
- RTX 3070 8 GB Q8_0 42K context · ~101 tok/s
- RTX 3070 Ti 8 GB Q8_0 42K context · ~138 tok/s
- RTX 4060 8 GB Q8_0 42K context · ~62 tok/s
- RTX 4060 Ti 8 GB Q8_0 42K context · ~65 tok/s
- RTX 5050 8 GB Q8_0 42K context · ~72 tok/s
- RTX 5060 8 GB Q8_0 42K context · ~101 tok/s
- RTX 5060 Ti 8 GB Q8_0 42K context · ~101 tok/s
- T1000 8GB 8 GB Q8_0 42K context · ~36 tok/s
- RTX 3080 10 GB fp16 33K context · ~102 tok/s
- GTX 1080 Ti 11 GB fp16 42K context · ~65 tok/s
- RTX 2080 Ti 11 GB fp16 42K context · ~83 tok/s
- RTX 2060 12GB 12 GB fp16 52K context · ~45 tok/s
- RTX 3060 12 GB fp16 52K context · ~48 tok/s
- RTX 3080 12GB 12 GB fp16 52K context · ~123 tok/s
- RTX 3080 Ti 12 GB fp16 52K context · ~123 tok/s
- RTX 4070 12 GB fp16 52K context · ~68 tok/s
- RTX 4070 SUPER 12 GB fp16 52K context · ~68 tok/s
- RTX 4070 Ti 12 GB fp16 52K context · ~68 tok/s
- RTX 5070 12 GB fp16 52K context · ~90 tok/s
- RTX A2000 12GB 12 GB fp16 52K context · ~39 tok/s
- RTX 4060 Ti 16GB 16 GB fp16 89K context · ~39 tok/s
- RTX 4070 Ti SUPER 16 GB fp16 89K context · ~90 tok/s
- RTX 4080 16 GB fp16 89K context · ~97 tok/s
- RTX 4080 SUPER 16 GB fp16 89K context · ~99 tok/s
- RTX 5060 Ti 16GB 16 GB fp16 89K context · ~60 tok/s
- RTX 5070 Ti 16 GB fp16 89K context · ~121 tok/s
- RTX 5080 16 GB fp16 89K context · ~129 tok/s
- RTX 2000 Ada 16 GB fp16 89K context · ~30 tok/s
- RTX A4000 16 GB fp16 89K context · ~60 tok/s
- RTX PRO 2000 Blackwell 16 GB fp16 89K context · ~39 tok/s
- Tesla T4 16 GB fp16 89K context · ~43 tok/s
- Tesla V100 16GB 16 GB fp16 89K context · ~121 tok/s
- RTX 4000 Ada 20 GB fp16 126K context · ~48 tok/s
- RTX 3090 24 GB fp16 128K context · ~126 tok/s
- RTX 3090 Ti 24 GB fp16 128K context · ~136 tok/s
- RTX 4090 24 GB fp16 128K context · ~136 tok/s
- RTX 4500 Ada 24 GB fp16 128K context · ~58 tok/s
- RTX A5000 24 GB fp16 128K context · ~103 tok/s
- RTX PRO 4000 Blackwell 24 GB fp16 128K context · ~90 tok/s
- RTX Spark 24 GB fp16 128K context · ~40 tok/s
- Tesla P40 24 GB fp16 128K context · ~47 tok/s
- RTX 5090 32 GB fp16 128K context · ~241 tok/s
- RTX 5000 Ada 32 GB fp16 128K context · ~78 tok/s
- RTX PRO 4500 Blackwell 32 GB fp16 128K context · ~121 tok/s
- Tesla V100 32GB 32 GB fp16 128K context · ~121 tok/s
- L40S 48 GB fp16 128K context · ~116 tok/s
- RTX 6000 Ada 48 GB fp16 128K context · ~129 tok/s
- RTX A6000 48 GB fp16 128K context · ~103 tok/s
- RTX PRO 5000 Blackwell 48 GB fp16 128K context · ~181 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB fp16 128K context · ~181 tok/s
- RTX PRO 5500 Blackwell 84 GB fp16 128K context · ~188 tok/s
- RTX PRO 6000 Blackwell 96 GB fp16 128K context · ~241 tok/s
- DGX Spark (GB 10) 128 GB fp16 128K context · ~37 tok/s
AMD
- Radeon RX 5600 XT 6 GB Q8_0 24K context · ~65 tok/s
- Radeon RX 5700 8 GB Q8_0 42K context · ~101 tok/s
- Radeon RX 5700 XT 8 GB Q8_0 42K context · ~101 tok/s
- Radeon RX 6600 8 GB Q8_0 42K context · ~51 tok/s
- Radeon RX 6600 XT 8 GB Q8_0 42K context · ~58 tok/s
- Radeon RX 6650 XT 8 GB Q8_0 42K context · ~63 tok/s
- Radeon RX 7600 8 GB Q8_0 42K context · ~65 tok/s
- Radeon RX 9060 8 GB Q8_0 42K context · ~65 tok/s
- Radeon RX 9060 XT 8 GB Q8_0 42K context · ~72 tok/s
- Radeon RX 6700 XT 12 GB fp16 52K context · ~52 tok/s
- Radeon RX 6750 XT 12 GB fp16 52K context · ~58 tok/s
- Radeon RX 7700 XT 12 GB fp16 52K context · ~58 tok/s
- Radeon RX 9070 GRE 12 GB fp16 52K context · ~58 tok/s
- AMD BC-250 16 GB fp16 89K context · ~60 tok/s
- Radeon RX 6800 16 GB fp16 89K context · ~69 tok/s
- Radeon RX 6800 XT 16 GB fp16 89K context · ~69 tok/s
- Radeon RX 6900 XT 16 GB fp16 89K context · ~69 tok/s
- Radeon RX 6950 XT 16 GB fp16 89K context · ~78 tok/s
- Radeon RX 7600 XT 16 GB fp16 89K context · ~39 tok/s
- Radeon RX 7800 XT 16 GB fp16 89K context · ~84 tok/s
- Radeon RX 7900 GRE 16 GB fp16 89K context · ~78 tok/s
- Radeon RX 9060 XT 16GB 16 GB fp16 89K context · ~43 tok/s
- Radeon RX 9070 16 GB fp16 89K context · ~86 tok/s
- Radeon RX 9070 XT 16 GB fp16 89K context · ~86 tok/s
- Radeon RX 7900 XT 20 GB fp16 126K context · ~108 tok/s
- Radeon RX 7900 XTX 24 GB fp16 128K context · ~129 tok/s
- Radeon AI PRO R9700 32 GB fp16 128K context · ~86 tok/s
- Ryzen AI Max (Strix Halo) 32 GB fp16 128K context · ~34 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB fp16 128K context · ~37 tok/s
Intel
- Arc A580 8 GB Q8_0 42K context · ~116 tok/s
- Arc A750 8 GB Q8_0 42K context · ~116 tok/s
- Arc A770 8GB 8 GB Q8_0 42K context · ~116 tok/s
- Arc B570 10 GB fp16 33K context · ~51 tok/s
- Arc B580 12 GB fp16 52K context · ~61 tok/s
- Arc A770 16GB 16 GB fp16 89K context · ~75 tok/s
- Arc Pro B65 32 GB fp16 128K context · ~82 tok/s
- Arc Pro B70 32 GB fp16 128K context · ~82 tok/s
Apple
- M1 8 GB Q8_0 42K context · ~15 tok/s
- M2 8 GB Q8_0 42K context · ~23 tok/s
- M3 8 GB Q8_0 42K context · ~23 tok/s
- M1 Pro 16 GB fp16 89K context · ~27 tok/s
- M2 Pro 16 GB fp16 89K context · ~27 tok/s
- M4 16 GB fp16 89K context · ~16 tok/s
- M5 16 GB fp16 89K context · ~21 tok/s
- M6 16 GB fp16 89K context · ~21 tok/s
- M3 Pro 18 GB fp16 108K context · ~20 tok/s
- M4 Pro 24 GB fp16 128K context · ~37 tok/s
- M5 Pro 24 GB fp16 128K context · ~41 tok/s
- M1 Max 32 GB fp16 128K context · ~54 tok/s
- M2 Max 32 GB fp16 128K context · ~54 tok/s
- M3 Max 36 GB fp16 128K context · ~40 tok/s
- M4 Max 36 GB fp16 128K context · ~55 tok/s
- M5 Max 36 GB fp16 128K context · ~62 tok/s
- M1 Ultra 64 GB fp16 128K context · ~108 tok/s
- M2 Ultra 64 GB fp16 128K context · ~108 tok/s
- M3 Ultra 96 GB fp16 128K context · ~110 tok/s
- M5 Ultra 96 GB fp16 128K context · ~162 tok/s