Models · chat and writing
Llama 3.2 3B: VRAM requirements and GPUs that run it
Best 3B general model. Solid instruction following for its size. At Q4 with 8K context it needs about 3.0 GB, so you want a GPU with at least 4 GB of VRAM or a Mac with 8 GB of unified memory.
Memory needed by quantization
| Quant | Weights | Total at 8K | Total at 32K |
|---|---|---|---|
| Q4_K_M | 2.1 GB | 3.0 GB | 5.6 GB |
| Q5_K_M | 2.5 GB | 3.4 GB | 6.1 GB |
| Q8_0 | 3.7 GB | 4.6 GB | 7.3 GB |
| FP16 | 7.0 GB | 7.9 GB | 11 GB |
Totals include about 10% runtime overhead and the KV cache. Longer conversations cost real memory: going from 8K to 32K context adds about 2.7 GB at any quantization. Most people should download Q4_K_M; Q5 or Q8 only make sense if you have spare memory. Read which quantization to pick.
Which GPUs can run Llama 3.2 3B?
40 of the 40 cards we track run it fully in VRAM at 8K context. The cheapest entry point is a 4 GB card such as the GTX 1650, and the fastest is the RTX 5090 at about 340 tokens per second.
| GPU | VRAM | Status | Quant | Speed |
|---|---|---|---|---|
| RTX 5090 | 32 GB | Runs well | Q8_0 | ~340 tok/s |
| RTX 4090 | 24 GB | Runs well | Q8_0 | ~190 tok/s |
| RTX 5080 | 16 GB | Runs well | Q8_0 | ~180 tok/s |
| RX 7900 XTX | 24 GB | Runs well | Q8_0 | ~180 tok/s |
| RTX 3090 / 3090 Ti | 24 GB | Runs well | Q8_0 | ~180 tok/s |
| RTX 5070 Ti | 16 GB | Runs well | Q8_0 | ~170 tok/s |
| RX 7900 XT | 20 GB | Runs well | Q8_0 | ~150 tok/s |
| RTX 3080 | 10 GB | Runs well | Q8_0 | ~150 tok/s |
| RTX 4080 / 4080 Super | 16 GB | Runs well | Q8_0 | ~140 tok/s |
| RTX 5070 | 12 GB | Runs well | Q8_0 | ~130 tok/s |
| RTX 4070 Ti Super | 16 GB | Runs well | Q8_0 | ~130 tok/s |
| RX 9070 XT | 16 GB | Runs well | Q8_0 | ~120 tok/s |
| RX 7800 XT | 16 GB | Runs well | Q8_0 | ~120 tok/s |
| RTX 4090 Laptop | 16 GB | Runs well | Q8_0 | ~110 tok/s |
| Arc A770 | 16 GB | Runs well | Q8_0 | ~110 tok/s |
| Arc A750 | 8 GB | Runs well | Q8_0 | ~98 tok/s |
| RX 6800 XT | 16 GB | Runs well | Q8_0 | ~98 tok/s |
| RTX 4070 / 4070 Super | 12 GB | Runs well | Q8_0 | ~97 tok/s |
| GTX 1080 Ti | 11 GB | Runs well | Q8_0 | ~93 tok/s |
| Arc B580 | 12 GB | Runs well | Q8_0 | ~87 tok/s |
| RTX 5060 Ti | 8 GB | Runs well | Q8_0 | ~86 tok/s |
| RTX 5060 | 8 GB | Runs well | Q8_0 | ~86 tok/s |
| RTX 3070 / 3070 Ti | 8 GB | Runs well | Q8_0 | ~86 tok/s |
| RTX 3060 Ti | 8 GB | Runs well | Q8_0 | ~86 tok/s |
| RTX 5060 Ti | 16 GB | Runs well | Q8_0 | ~86 tok/s |
| RTX 4080 Laptop | 12 GB | Runs well | Q8_0 | ~83 tok/s |
| RX 7700 XT | 12 GB | Runs well | Q8_0 | ~83 tok/s |
| RX 6700 XT | 12 GB | Runs well | Q8_0 | ~74 tok/s |
| RTX 3060 | 12 GB | Runs well | Q8_0 | ~69 tok/s |
| GTX 1660 / 1660 Super | 6 GB | Runs well | Q8_0 | ~64 tok/s |
| RTX 3060 Laptop | 6 GB | Runs well | Q8_0 | ~64 tok/s |
| RTX 4060 Ti | 8 GB | Runs well | Q8_0 | ~55 tok/s |
| RX 7600 | 8 GB | Runs well | Q8_0 | ~55 tok/s |
| RTX 4060 Ti | 16 GB | Runs well | Q8_0 | ~55 tok/s |
| RTX 4060 | 8 GB | Runs well | Q8_0 | ~52 tok/s |
| RTX 4070 Laptop | 8 GB | Runs well | Q8_0 | ~49 tok/s |
| RTX 4060 Laptop | 8 GB | Runs well | Q8_0 | ~49 tok/s |
| RTX 3050 | 8 GB | Runs well | Q8_0 | ~43 tok/s |
| RTX 4050 Laptop | 6 GB | Runs well | Q8_0 | ~37 tok/s |
| GTX 1650 | 4 GB | Runs well | Q5_K_M | ~37 tok/s |
Running it on a Mac
Apple Silicon Macs can give a model about 72% of their unified memory. Llama 3.2 3B fits from 8 GB (Apple M1, about 13 tokens per second). With a faster chip like the Apple M3 Ultra (96 GB) expect around 160 tokens per second. Compare the options in our Mac vs NVIDIA guide.
Similar size models: Phi-4 Mini 3.8B, Qwen3 4B, Gemma 3 4B.
Compare Llama 3.2 3B
Questions
How much VRAM does Llama 3.2 3B need?
About 3.0 GB at Q4_K_M with 8K context, 4.6 GB at Q8_0, and 7.9 GB at full FP16. In practice that means a card with at least 4 GB of VRAM for Q4.
Can Llama 3.2 3B run on an 8 GB GPU?
Yes, at Q8_0, using about 4.6 GB.
Can I run Llama 3.2 3B on a Mac?
Yes. The smallest configuration that runs it fully is Apple M1 with 8 GB of unified memory, at about 13 tokens per second. More memory bandwidth (Pro, Max, Ultra chips) means faster replies.
Estimates based on model size, architecture and memory bandwidth. Real numbers vary with software, drivers and context length.