Llama-3.2-3B
by meta-llama
3.2B params · text-generation · 704 likes · 1.4M downloads
Llama-3.2-3B is a 3.2B parameter model. At Q4 quantization it requires 2GB of VRAM. It runs comfortably on GeForce RTX 4090 (408 tok/s), GeForce RTX 5090 (612 tok/s), M4 Max 128GB (149 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 408 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 612 tok/s | comfortable |
| M4 Max 128GB | 128GB | 149 tok/s | comfortable |
| M4 Pro 48GB | 48GB | 74 tok/s | comfortable |
| M4 Pro 24GB | 24GB | 74 tok/s | comfortable |
| A100 PCIe 80 GB | 80GB | 748 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 1505 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 361 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 298 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 290 tok/s | comfortable |