Can I run Llama 3.2 3B Instruct locally?
Llama 3.2 3B Instruct by meta-llama has 3.2B parameters. At 4-bit with an 8K context it needs about 3.4 GB of GPU memory. It fits on 65 of 65 GPUs and Macs we track. See all models that fit in 8 GB of VRAM.
Memory needed by quant and context length
GB of GPU memory = weights + KV cache (FP16) + runtime overhead. Weights use real file sizes from unsloth/Llama-3.2-3B-Instruct-GGUF where available (marked •), otherwise parameter count × bits per weight. How we calculate.
| Quant | Quality | 4K ctx | 8K ctx | 32K ctx | 128K ctx |
|---|---|---|---|---|---|
| 2-bit (Q2_K) • | rough | 2.4 GB | 2.8 GB | 5.4 GB | 15.9 GB |
| 3-bit (Q3_K_M) • | usable | 2.7 GB | 3.1 GB | 5.7 GB | 16.2 GB |
| 4-bit (Q4_K_M) • | good, most popular | 3.0 GB | 3.4 GB | 6.0 GB | 16.5 GB |
| 5-bit (Q5_K_M) • | very good | 3.2 GB | 3.7 GB | 6.3 GB | 16.8 GB |
| 6-bit (Q6_K) • | near lossless | 3.5 GB | 4.0 GB | 6.6 GB | 17.1 GB |
| 8-bit (Q8_0) • | effectively lossless | 4.3 GB | 4.7 GB | 7.3 GB | 17.8 GB |
| 16-bit (F16/BF16) • | original weights | 13.1 GB | 13.5 GB | 16.1 GB | 26.6 GB |
Which GPUs and Macs run Llama 3.2 3B Instruct?
Best quality quant that fits at 8K context. "Tight" = within 10% of usable memory. Speed is a ceiling from memory bandwidth (real speed is typically 50–80% of it).
| Hardware | Memory | Fits? | Best quant | Needs | Max tok/s (est.) | |
|---|---|---|---|---|---|---|
| RTX 3060 12GB | 12 GB | Fits | Q8_0 | 4.7 GB | 105 | details |
| RTX 4060 Ti 16GB | 16 GB | Fits | F16 | 13.5 GB | 22 | details |
| RTX 3090 | 24 GB | Fits | F16 | 13.5 GB | 73 | details |
| RTX 4090 | 24 GB | Fits | F16 | 13.5 GB | 78 | details |
| RTX 5090 | 32 GB | Fits | F16 | 13.5 GB | 139 | details |
| RTX 5070 Ti | 16 GB | Fits | F16 | 13.5 GB | 70 | details |
| RX 7900 XTX | 24 GB | Fits | F16 | 13.5 GB | 75 | details |
| Mac M4 Pro 48GB | 48 GB | Fits | F16 | 13.5 GB | 21 | details |
| Mac M4 Max 128GB | 128 GB | Fits | F16 | 13.5 GB | 42 | details |
| Mac M3 Ultra 512GB | 512 GB | Fits | F16 | 13.5 GB | 64 | details |
| DGX Spark | 128 GB | Fits | F16 | 13.5 GB | 21 | details |
| RTX 4070 | 12 GB | Fits | Q8_0 | 4.7 GB | 147 | details |
| RTX 5060 Ti 16GB | 16 GB | Fits | F16 | 13.5 GB | 35 | details |
| RTX 4060 | 8 GB | Fits | Q8_0 | 4.7 GB | 79 | details |
| Arc B580 | 12 GB | Fits | Q8_0 | 4.7 GB | 133 | details |
| RTX 3060 Ti | 8 GB | Fits | Q8_0 | 4.7 GB | 131 | details |
| RTX 3070 | 8 GB | Fits | Q8_0 | 4.7 GB | 131 | details |
| RTX 3080 10GB | 10 GB | Fits | Q8_0 | 4.7 GB | 222 | details |
| RTX 3080 12GB | 12 GB | Fits | Q8_0 | 4.7 GB | 267 | details |
| RTX 3090 Ti | 24 GB | Fits | F16 | 13.5 GB | 78 | details |
| RTX 4060 Ti 8GB | 8 GB | Fits | Q8_0 | 4.7 GB | 84 | details |
| RTX 4070 Super | 12 GB | Fits | Q8_0 | 4.7 GB | 147 | details |
| RTX 4070 Ti | 12 GB | Fits | Q8_0 | 4.7 GB | 147 | details |
| RTX 4070 Ti Super | 16 GB | Fits | F16 | 13.5 GB | 52 | details |
| RTX 4080 | 16 GB | Fits | F16 | 13.5 GB | 56 | details |
| RTX 4080 Super | 16 GB | Fits | F16 | 13.5 GB | 57 | details |
| RTX 5060 | 8 GB | Fits | Q8_0 | 4.7 GB | 131 | details |
| RTX 5070 | 12 GB | Fits | Q8_0 | 4.7 GB | 196 | details |
| RTX 5080 | 16 GB | Fits | F16 | 13.5 GB | 75 | details |
| 2x RTX 3090 | 2×24 GB | Fits | F16 | 13.5 GB | 73 | details |
| 2x RTX 4090 | 2×24 GB | Fits | F16 | 13.5 GB | 78 | details |
| 2x RTX 5090 | 2×32 GB | Fits | F16 | 13.5 GB | 139 | details |
| RTX A6000 | 48 GB | Fits | F16 | 13.5 GB | 60 | details |
| RTX 6000 Ada | 48 GB | Fits | F16 | 13.5 GB | 75 | details |
| RTX PRO 6000 Blackwell | 96 GB | Fits | F16 | 13.5 GB | 139 | details |
| L40S | 48 GB | Fits | F16 | 13.5 GB | 67 | details |
| A100 40GB | 40 GB | Fits | F16 | 13.5 GB | 121 | details |
| A100 80GB | 80 GB | Fits | F16 | 13.5 GB | 158 | details |
| H100 PCIe | 80 GB | Fits | F16 | 13.5 GB | 155 | details |
| H100 SXM | 80 GB | Fits | F16 | 13.5 GB | 260 | details |
| H200 | 141 GB | Fits | F16 | 13.5 GB | 373 | details |
| RX 7600 | 8 GB | Fits | Q8_0 | 4.7 GB | 84 | details |
| RX 7700 XT | 12 GB | Fits | Q8_0 | 4.7 GB | 126 | details |
| RX 7800 XT | 16 GB | Fits | F16 | 13.5 GB | 48 | details |
| RX 7900 GRE | 16 GB | Fits | F16 | 13.5 GB | 45 | details |
| RX 7900 XT | 20 GB | Fits | F16 | 13.5 GB | 62 | details |
| RX 9070 | 16 GB | Fits | F16 | 13.5 GB | 50 | details |
| RX 9070 XT | 16 GB | Fits | F16 | 13.5 GB | 50 | details |
| Ryzen AI Max+ 395 128GB | 128 GB | Fits | F16 | 13.5 GB | 20 | details |
| Arc A770 16GB | 16 GB | Fits | F16 | 13.5 GB | 44 | details |
| Mac M1 16GB | 16 GB | Fits | Q8_0 | 4.7 GB | 20 | details |
| Mac M2 24GB | 24 GB | Fits | F16 | 13.5 GB | 7.8 | details |
| Mac M3 24GB | 24 GB | Fits | F16 | 13.5 GB | 7.8 | details |
| Mac M4 16GB | 16 GB | Fits | Q8_0 | 4.7 GB | 35 | details |
| Mac M4 32GB | 32 GB | Fits | F16 | 13.5 GB | 9.3 | details |
| Mac M1 Pro 32GB | 32 GB | Fits | F16 | 13.5 GB | 16 | details |
| Mac M3 Pro 36GB | 36 GB | Fits | F16 | 13.5 GB | 12 | details |
| Mac M4 Pro 24GB | 24 GB | Fits | F16 | 13.5 GB | 21 | details |
| Mac M4 Pro 64GB | 64 GB | Fits | F16 | 13.5 GB | 21 | details |
| Mac M1 Max 64GB | 64 GB | Fits | F16 | 13.5 GB | 31 | details |
| Mac M2 Max 96GB | 96 GB | Fits | F16 | 13.5 GB | 31 | details |
| Mac M3 Max 128GB | 128 GB | Fits | F16 | 13.5 GB | 31 | details |
| Mac M4 Max 64GB | 64 GB | Fits | F16 | 13.5 GB | 42 | details |
| Mac M2 Ultra 192GB | 192 GB | Fits | F16 | 13.5 GB | 62 | details |
| Mac M3 Ultra 256GB | 256 GB | Fits | F16 | 13.5 GB | 64 | details |
Prefer an API?
Meta: Llama 3.2 3B Instruct is on OpenRouter at $0.0509 per 1M input tokens and $0.335 per 1M output tokens. Price details and history →
Questions
How much VRAM does Llama 3.2 3B Instruct need?
About 3.4 GB at 4-bit (Q4_K_M) with an 8K context: 1.9 GB of weights (real GGUF file size), 0.9 GB of KV cache and 0.6 GB of runtime overhead. At 8-bit it needs about 4.7 GB.
What is the smallest GPU that runs Llama 3.2 3B Instruct?
At Q4_K_M and 8K context, the smallest single consumer GPU in our table that fits it is the RTX 4060 (8 GB).
Can I run Llama 3.2 3B Instruct on a Mac?
Yes. At Q4_K_M with 8K context it fits on a Mac M1 16GB or larger, using the default macOS GPU memory limit.
Similar models
Same family or a similar parameter count, with the VRAM each needs at Q4_K_M and 8K context.
Llama 3.2 3B
3.2B · 3.6 GB at 4-bitLlama 3.2 1B
1.2B · 1.8 GB at 4-bitLlama 3.2 1B Instruct
1.2B · 1.6 GB at 4-bitHermes 3 Llama 3.2 3B
3.2B · 3.4 GB at 4-bitgranite 3.1 3b a800m instruct
3.3B · 2.9 GB at 4-bitd1 3B
3.1B · 2.3 GB at 4-bitModel facts
| Hugging Face repo | meta-llama/Llama-3.2-3B-Instruct |
| Parameters | 3,212,749,824 (3.2B) |
| Layers / hidden size | 28 / 3072 |
| Attention | GQA; 24 heads, 8 KV heads, head dim 128 |
| KV cache per token (FP16) | 112.0 KiB |
| Max context | 131,072 tokens |
| License | llama3.2 (gated: accept terms on Hugging Face) |
| Released on HF | Sep 18, 2024 |
| Downloads (30d) / likes | 824,000 / 3,027 |
Sources: Hugging Face model API · config.json (unsloth/Llama-3.2-3B-Instruct-GGUF) · GGUF files · GPU specs from vendor pages (list). Data fetched Nov 17, 2026. Spot an error? Report it.