HuggingFace · SmolLM2 · 1.7B parameters
SmolLM2 1.7B VRAM requirements
SmolLM2 1.7B has 24 layers and uses full multi-head attention — no GQA, so the cache is large. At Q4_K_M the weights come to 1.0 GB, and the best quantisation that fits a 24 GB card is FP16 / BF16.
Runs comfortably
3.3 GB of 21.8 GB · 15%022 GB
Weights
1.0 GB
KV cache
1.5 GB
Runtime overhead
0.8 GB
SmolLM2 1.7B at Q4_K_M leaves 18.5 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.
Generation334tok/s
Prompt processing20263tok/s
Max context8Ktokens
KV per 1K tokens0GB
Every quantisation of SmolLM2 1.7B on a RTX 4090
Highlighted row is the highest quality that still fits at 8K context.
| Quantisation | bpw | Weights | Total | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 3.2 GB | 5.5 GB | Runs comfortably | 8K | 150 |
| INT8 / W8A8 | 8.50 | 1.7 GB | 4.0 GB | Runs comfortably | 8K | 241 |
| Q8_0 (GGUF) | 8.50 | 1.7 GB | 4.0 GB | Runs comfortably | 8K | 241 |
| FP8 (E4M3) | 8.00 | 1.6 GB | 3.9 GB | Runs comfortably | 8K | 249 |
| Q6_K | 6.56 | 1.3 GB | 3.6 GB | Runs comfortably | 8K | 282 |
| Q5_K_M | 5.67 | 1.1 GB | 3.4 GB | Runs comfortably | 8K | 308 |
| Q5_K_S | 5.52 | 1.1 GB | 3.4 GB | Runs comfortably | 8K | 312 |
| AWQ 4-bit | 4.25 | 1.0 GB | 3.3 GB | Runs comfortably | 8K | 332 |
| GPTQ 4-bit | 4.25 | 1.0 GB | 3.3 GB | Runs comfortably | 8K | 332 |
| MXFP4 | 4.25 | 1.0 GB | 3.3 GB | Runs comfortably | 8K | 332 |
| Q4_K_M | 4.85 | 1.0 GB | 3.3 GB | Runs comfortably | 8K | 334 |
| Q4_K_S | 4.58 | 0.9 GB | 3.2 GB | Runs comfortably | 8K | 344 |
| Q4_0 | 4.55 | 0.9 GB | 3.2 GB | Runs comfortably | 8K | 345 |
| IQ4_XS | 4.25 | 0.9 GB | 3.1 GB | Runs comfortably | 8K | 356 |
| Q3_K_M | 3.91 | 0.8 GB | 3.1 GB | Runs comfortably | 8K | 370 |
| IQ3_M | 3.70 | 0.8 GB | 3.0 GB | Runs comfortably | 8K | 379 |
| IQ3_XXS | 3.06 | 0.6 GB | 2.9 GB | Runs comfortably | 8K | 410 |
| Q2_K | 2.63 | 0.6 GB | 2.8 GB | Runs comfortably | 8K | 434 |
| IQ2_XXS | 2.06 | 0.5 GB | 2.7 GB | Runs comfortably | 8K | 470 |
| IQ1_M | 1.75 | 0.4 GB | 2.7 GB | Runs comfortably | 8K | 492 |
SmolLM2 1.7B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| H100 SXM 80GB | 80 | 3350 | Runs comfortably | 8K | 1138 |
| A100 80GB | 80 | 2039 | Runs comfortably | 8K | 669 |
| RTX 5090 | 32 | 1792 | Runs comfortably | 8K | 625 |
| RTX 5080 | 16 | 960 | Runs comfortably | 8K | 350 |
| RTX 4090 | 24 | 1008 | Runs comfortably | 8K | 334 |
| RTX 5070 Ti | 16 | 896 | Runs comfortably | 8K | 327 |
| RTX 3090 | 24 | 936 | Runs comfortably | 8K | 324 |
| Radeon RX 7900 XTX | 24 | 960 | Runs comfortably | 8K | 303 |
| L40S | 48 | 864 | Runs comfortably | 8K | 288 |
| RTX A6000 | 48 | 768 | Runs comfortably | 8K | 268 |
| RTX 3080 10GB | 10 | 760 | Runs comfortably | 8K | 266 |
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 8K | 264 |
| RTX 5070 | 12 | 672 | Runs comfortably | 8K | 249 |
| RTX 4080 Super | 16 | 736 | Runs comfortably | 8K | 247 |
| RTX 4070 Ti Super | 16 | 672 | Runs comfortably | 8K | 226 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 8K | 198 |
| RTX 4070 Super | 12 | 504 | Runs comfortably | 8K | 171 |
| RTX 4070 | 12 | 504 | Runs comfortably | 8K | 171 |
| RTX 5060 Ti 16GB | 16 | 448 | Runs comfortably | 8K | 168 |
| Arc B580 | 12 | 456 | Runs comfortably | 8K | 132 |
| RTX 3060 12GB | 12 | 360 | Runs comfortably | 8K | 128 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 8K | 103 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Runs comfortably | 8K | 100 |
| RTX 4060 Ti 16GB | 16 | 288 | Runs comfortably | 8K | 98.8 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 8K | 77.2 |
Architecture
| Parameters | 1.7B |
| Layers | 24 |
| Hidden size | 2048 |
| Attention heads / KV heads | 32 / 32 |
| Head dimension | 64 |
| Vocabulary | 49,152 |
| Trained context | 8K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | HuggingFaceTB/SmolLM2-1.7B-Instruct |
The HuggingFace family
SmolLM2 comes from Hugging Face's own team and targets the small end — 1.7B parameters, a 49k vocabulary and full multi-head attention. It runs on nearly anything.
huggingface.co/HuggingFaceTB · github.com/huggingface/smollm