SmolLM2-1.7B-Instruct
by HuggingFaceTB
1.8B params · text-generation · 720 likes · 116.3k downloads
SmolLM2-1.7B-Instruct is a 1.8B parameter model. At Q4 quantization it requires 1GB of VRAM. It runs comfortably on GeForce RTX 4090 (724 tok/s), GeForce RTX 5090 (1086 tok/s), M4 Max 128GB (265 tok/s).
GPU compatibility
| GPU | VRAM | Q4 Decode | Verdict |
|---|---|---|---|
| GeForce RTX 4090 | 24GB | 724 tok/s | comfortable |
| GeForce RTX 5090 | 32GB | 1086 tok/s | comfortable |
| M4 Max 128GB | 128GB | 265 tok/s | comfortable |
| M4 Pro 48GB | 48GB | 132 tok/s | comfortable |
| M4 Pro 24GB | 24GB | 132 tok/s | comfortable |
| A100 PCIe 80 GB | 80GB | 1327 tok/s | comfortable |
| H100 SXM5 80 GB | 80GB | 2670 tok/s | comfortable |
| GeForce RTX 3090 | 24GB | 640 tok/s | comfortable |
| Radeon RX 7900 XTX | 24GB | 529 tok/s | comfortable |
| GeForce RTX 4080 | 16GB | 514 tok/s | comfortable |