Written by Jakub Rusinowski · Last updated January 20, 2025
These figures are for DeepSeek-R1-Distill-Qwen-32B, a distill of Qwen2.5-32B — not the full DeepSeek R1. The full DeepSeek R1 (671B) needs about 405 GB of weights at Q4_K_M and is a different model.
Yes, but it is tight
Yes, but it is tight — DeepSeek R1 Distill Qwen 32B at Q4_K_M needs about 22.3 GB of the 24 GB usable on RTX 3090 Desktop (24 GB VRAM, 64 GB RAM), leaving only ~1.7 GB before the runtime starts swapping. Expect ~32 tok/s (estimated), with room for about 8,192 tokens of context.
Confidence: medium · Recommended quantization: Q4_K_M · Estimated speed: ~32 tok/s
| Usable memory for models | 24 GB |
| Memory bandwidth | 936 GB/s |
| Form factor | Desktop |
| Operating system | Windows or Linux |
| Memory upgradeable | Yes |
| Quant | Memory needed | Fits 24 GB? | Max context | Est. speed | Download |
|---|---|---|---|---|---|
| F16 | 66.9 GB | ✗ No | — | — | 64 GB |
| Q8_0 | 36.9 GB | ✗ No | — | — | 34 GB |
| Q6_K | 29.2 GB | ✗ No | — | — | 26.2 GB |
| Q5_K_M | 25.6 GB | ✗ No | — | — | 22.7 GB |
| Q4_K_M | 22.3 GB | ✓ Yes | 8K | ~32 tok/s | 19.3 GB |
| Q3_K_M | 16.6 GB | ✓ Yes | 32K | ~43.1 tok/s | 13.6 GB |
| Q2_K | 13.5 GB | ✓ Yes | 32K | ~53.3 tok/s | 10.5 GB |
| Variant | Needs at Q4_K_M | Fits? | Est. speed |
|---|---|---|---|
| DeepSeek R1 (671B) | 406.5 GB | ✗ Too large | — |
| DeepSeek R1 Distill Qwen 32B | 22.3 GB | ✓ Fits | ~32 tok/s |
| DeepSeek R1 Distill Qwen 14B | 10.6 GB | ✓ Fits | ~65.6 tok/s |
| DeepSeek R1 Distill Llama 8B | 6.7 GB | ✓ Fits | ~101.1 tok/s |
Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving
Yes, but it is tight — DeepSeek R1 Distill Qwen 32B at Q4_K_M needs about 22.3 GB of the 24 GB usable on RTX 3090 Desktop (24 GB VRAM, 64 GB RAM), leaving only ~1.7 GB before the runtime starts swapping. Expect ~32 tok/s (estimated), with room for about 8,192 tokens of context.
Q4_K_M — it needs about 22.3 GB of the 24 GB available, downloads as roughly 19.3 GB, and runs at an estimated 32 tokens/sec with up to 8K of context.
Nothing binding — the model fits with headroom and generates at a usable speed on this hardware.
Ollama or llama.cpp (CUDA/ROCm) — vLLM if you need concurrent serving
← Can I Run It? | DeepSeek R1 model page | Check your hardware