Qwen · Qwen3 · 1.7B parameters
Qwen3 1.7B VRAM requirements
Qwen3 1.7B has 28 layers and uses grouped-query attention (8 KV heads). 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
2.7 GB of 21.8 GB · 12%Qwen3 1.7B at Q4_K_M leaves 19.1 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.
Every quantisation of Qwen3 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 | 4.9 GB | Runs comfortably | 32K | 162 |
| INT8 / W8A8 | 8.50 | 1.6 GB | 3.3 GB | Runs comfortably | 32K | 278 |
| Q8_0 (GGUF) | 8.50 | 1.6 GB | 3.3 GB | Runs comfortably | 32K | 278 |
| FP8 (E4M3) | 8.00 | 1.6 GB | 3.3 GB | Runs comfortably | 32K | 282 |
| Q6_K | 6.56 | 1.3 GB | 3.0 GB | Runs comfortably | 32K | 326 |
| AWQ 4-bit | 4.25 | 1.3 GB | 2.9 GB | Runs comfortably | 32K | 333 |
| GPTQ 4-bit | 4.25 | 1.3 GB | 2.9 GB | Runs comfortably | 32K | 333 |
| MXFP4 | 4.25 | 1.3 GB | 2.9 GB | Runs comfortably | 32K | 333 |
| Q5_K_M | 5.67 | 1.1 GB | 2.8 GB | Runs comfortably | 32K | 361 |
| Q5_K_S | 5.52 | 1.1 GB | 2.8 GB | Runs comfortably | 32K | 368 |
| Q4_K_M | 4.85 | 1.0 GB | 2.7 GB | Runs comfortably | 32K | 394 |
| Q4_K_S | 4.58 | 1.0 GB | 2.6 GB | Runs comfortably | 32K | 406 |
| Q4_0 | 4.55 | 0.9 GB | 2.6 GB | Runs comfortably | 32K | 407 |
| IQ4_XS | 4.25 | 0.9 GB | 2.6 GB | Runs comfortably | 32K | 421 |
| Q3_K_M | 3.91 | 0.8 GB | 2.5 GB | Runs comfortably | 32K | 438 |
| IQ3_M | 3.70 | 0.8 GB | 2.5 GB | Runs comfortably | 32K | 449 |
| IQ3_XXS | 3.06 | 0.7 GB | 2.4 GB | Runs comfortably | 32K | 487 |
| Q2_K | 2.63 | 0.6 GB | 2.3 GB | Runs comfortably | 32K | 517 |
| IQ2_XXS | 2.06 | 0.5 GB | 2.2 GB | Runs comfortably | 32K | 562 |
| IQ1_M | 1.75 | 0.5 GB | 2.1 GB | Runs comfortably | 32K | 590 |
Qwen3 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 | 32K | 1285 |
| A100 80GB | 80 | 2039 | Runs comfortably | 32K | 775 |
| RTX 5090 | 32 | 1792 | Runs comfortably | 32K | 726 |
| RTX 5080 | 16 | 960 | Runs comfortably | 32K | 412 |
| RTX 4090 | 24 | 1008 | Runs comfortably | 32K | 394 |
| RTX 5070 Ti | 16 | 896 | Runs comfortably | 32K | 387 |
| RTX 3090 | 24 | 936 | Runs comfortably | 32K | 383 |
| Radeon RX 7900 XTX | 24 | 960 | Runs comfortably | 32K | 358 |
| L40S | 48 | 864 | Runs comfortably | 32K | 341 |
| RTX A6000 | 48 | 768 | Runs comfortably | 32K | 318 |
| RTX 3080 10GB | 10 | 760 | Runs comfortably | 32K | 315 |
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 32K | 313 |
| RTX 5070 | 12 | 672 | Runs comfortably | 32K | 295 |
| RTX 4080 Super | 16 | 736 | Runs comfortably | 32K | 293 |
| RTX 4070 Ti Super | 16 | 672 | Runs comfortably | 32K | 269 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 32K | 236 |
| RTX 4070 Super | 12 | 504 | Runs comfortably | 32K | 204 |
| RTX 4070 | 12 | 504 | Runs comfortably | 32K | 204 |
| RTX 5060 Ti 16GB | 16 | 448 | Runs comfortably | 32K | 200 |
| Arc B580 | 12 | 456 | Runs comfortably | 32K | 157 |
| RTX 3060 12GB | 12 | 360 | Runs comfortably | 32K | 153 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 32K | 123 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Runs comfortably | 32K | 120 |
| RTX 4060 Ti 16GB | 16 | 288 | Runs comfortably | 32K | 118 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 32K | 92.4 |
Architecture
| Parameters | 1.7B |
| Layers | 28 |
| Hidden size | 2048 |
| Attention heads / KV heads | 16 / 8 |
| Head dimension | 128 |
| Vocabulary | 151,936 |
| Trained context | 32K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | Qwen/Qwen3-1.7B |
The Qwen family
Alibaba's series, and the broadest size ladder available — Qwen3 runs from 0.6B to 32B dense, plus 30B-A3B and 235B-A22B as mixture-of-experts. The 152k vocabulary makes the embedding table a large share of a small model's file. Qwen2.5-Coder is the same architecture trained for code.