Llama 3.2 3B VRAM Requirements
Llama 3.2 3B needs about 3.5 GB of VRAM at Q4_K_M with an 8K context: 1.8 GB for weights, 0.9 GB for the KV cache and 0.8 GB of runtime overhead. At Q8_0 it needs 4.8 GB and at full FP16 precision 7.6 GB. The smallest GPU that runs it fully is the RTX 3050 8GB.
3.5 GB VRAM
Llama 3.2 3B at Q4_K_M with 8K context and the KV cache in FP16.
- Weights 1.8 GB
- KV cache 0.9 GB
- Runtime overhead 0.8 GB
Which GPUs can run it
- RTX 3060 12GBFits~94 tok/s
- RTX 4060 Ti 16GBFits~76 tok/s
- RTX 3090 24GBFits~235 tok/s
- Radeon RX 7900 XTX 24GBFits~206 tok/s
- RTX 4090 24GBFits~252 tok/s
- RTX 5090 32GBFits~425 tok/s
- RTX 6000 Ada 48GBFits~241 tok/s
- A100 80GBFits~476 tok/s
- H100 80GB SXMFits~720 tok/s
- Ryzen AI Max+ 395 128GBFits~62 tok/s
- Apple M4 Max 128GBFits~141 tok/s
- Apple M5 Max 128GBFits~158 tok/s
- RTX PRO 6000 Blackwell 96GBFits~425 tok/s
- NVIDIA DGX Spark 128GBFits~67 tok/s
- Apple M3 Ultra 512GBFits~208 tok/s
Llama 3.2 3B VRAM by quantization
Total memory with an FP16 KV cache. "Smallest tier" is the smallest standard VRAM size with a 6% safety margin at 8K.
| Quantization | Bits/weight | Weights | Total at 8K | Total at 32K | Smallest tier |
|---|---|---|---|---|---|
| Q4_K_M | 4.85 | 1.8 GB | 3.5 GB | 6.1 GB | 4 GB |
| Q5_K_M | 5.69 | 2.1 GB | 3.8 GB | 6.4 GB | 6 GB |
| Q6_K | 6.59 | 2.5 GB | 4.1 GB | 6.8 GB | 6 GB |
| Q8_0 | 8.50 | 3.2 GB | 4.8 GB | 7.5 GB | 6 GB |
| FP16 / BF16 | 16.00 | 6.0 GB | 7.6 GB | 10.3 GB | 10 GB |
| Q3_K_M | 3.91 | 1.5 GB | 3.1 GB | 5.8 GB | 4 GB |
| Q2_K | 3.00 | 1.1 GB | 2.8 GB | 5.4 GB | 4 GB |
VRAM by context length
At Q4_K_M. The KV cache grows linearly with context: 0.11 GB per 1,000 tokens in FP16.
| Context | KV cache (FP16) | KV cache (Q8_0) | Total, FP16 cache | Total, Q8_0 cache |
|---|---|---|---|---|
| 4K tokens | 0.4 GB | 0.2 GB | 3.0 GB | 2.8 GB |
| 8K tokens | 0.9 GB | 0.5 GB | 3.5 GB | 3.1 GB |
| 32K tokens | 3.5 GB | 1.9 GB | 6.1 GB | 4.5 GB |
| 128K tokens | 14.0 GB | 7.4 GB | 16.6 GB | 10.0 GB |
GPUs that can run Llama 3.2 3B
At Q4_K_M with 8K context. "Best quant" is the highest-quality format that fits fully; "max context" is at Q4_K_M.
| GPU | Usable memory | Status | Best quant | Max context | Speed |
|---|---|---|---|---|---|
| RTX 3050 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~59 tok/s |
| RTX 4060 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~72 tok/s |
| RTX 4060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~76 tok/s |
| Radeon RX 7600 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~64 tok/s |
| RTX 3060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~117 tok/s |
| RTX 3070 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~117 tok/s |
| RTX 5060 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~117 tok/s |
| RTX 5060 Ti 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~117 tok/s |
| RTX 3070 Ti 8GB | 8.0 GB | Fits | Q8_0 | 44K | ~156 tok/s |
| Intel Arc B570 10GB | 10.0 GB | Fits | FP16 / BF16 | 62K | ~77 tok/s |
| RTX 3080 10GB | 10.0 GB | Fits | FP16 / BF16 | 62K | ~193 tok/s |
| Apple M4 16GB | 10.7 GB | Fits | FP16 / BF16 | 68K | ~32 tok/s |
| RTX 2080 Ti 11GB | 11.0 GB | Fits | FP16 / BF16 | 70K | ~158 tok/s |
| RTX 2060 12GB | 12.0 GB | Fits | FP16 / BF16 | 79K | ~88 tok/s |
About Llama 3.2 3B
A compact Llama with the full 128K context window. Its KV cache grows faster than the 1B model's because it has 28 layers with 128-dimension heads, so long-context use can need more memory than the weights themselves.
| Spec | Value |
|---|---|
| Parameters | 3.21B |
| Layers | 28 |
| Hidden size | 3072 |
| Attention heads / KV heads | 24 / 8 |
| Head dimension | 128 |
| Vocabulary | 128,256 |
| Max context (used here) | 128K tokens |
| KV cache per token (FP16) | 112 KB |
| Released | 2024-09 |
| License | Llama 3.2 Community License |
Llama 3.2 3B VRAM: common questions
Is Llama 3.2 3B good for a 4GB or 6GB GPU?
Yes. At Q4_K_M it needs about 3.5 GB with an 8K context, which leaves room for a longer context on 6 GB cards.
How much VRAM does Llama 3.2 3B need?
About 3.5 GB at Q4_K_M with an 8K context, 4.8 GB at Q8_0 and 7.6 GB at FP16. Longer contexts add 0.11 GB per 1,000 tokens with an FP16 cache.
What is the cheapest GPU that can run Llama 3.2 3B?
The smallest listed device that fits it fully at Q4_K_M with 8K context is the RTX 3050 8GB (8.0 GB usable).