What GPU do you need to run QwQ 32B?
32.5B params 128K max context ReasoningTool use
The smallest GPU memory that runs QwQ 32B well is 24 GB (Q4_K_M, e.g. RTX Spark). 45 of 132 GPUs run it well.
MMLU 83
VRAM needed per quantization
| Quantization | Weights | 16K context | 32K context | Full 128K |
|---|---|---|---|---|
| Q4_K_M | 19.9 GB | 23.8 GB | 27.7 GB | 51.1 GB |
| Q8_0 | 34.9 GB | 38.8 GB | 42.7 GB | 66.1 GB |
Weights plus fp16 KV cache. Leave ~1 GB headroom for the inference engine.
Cheapest GPUs that can run QwQ 32B
Lowest launch MSRP among GPUs that run its smallest quantization (Q4_K_M) well, with at least 16K context.
As an Amazon Associate, I earn from qualifying purchases.
GPUs that can run QwQ 32B
Best quantization that fits on each GPU, with the context and speed it gets. For chips with several memory sizes, the smallest that works is shown.
NVIDIA
- RTX 3090 24 GB Q4_K_M 16K context · ~45 tok/s
- RTX 3090 Ti 24 GB Q4_K_M 16K context · ~48 tok/s
- RTX 4090 24 GB Q4_K_M 16K context · ~48 tok/s
- RTX 4500 Ada 24 GB Q4_K_M 16K context · ~21 tok/s
- RTX A5000 24 GB Q4_K_M 16K context · ~37 tok/s
- RTX PRO 4000 Blackwell 24 GB Q4_K_M 16K context · ~32 tok/s
- RTX Spark 24 GB Q4_K_M 16K context · ~14 tok/s
- Tesla P40 24 GB Q4_K_M 16K context · ~17 tok/s
- RTX 5090 32 GB Q4_K_M 49K context · ~86 tok/s
- RTX 5000 Ada 32 GB Q4_K_M 49K context · ~28 tok/s
- RTX PRO 4500 Blackwell 32 GB Q4_K_M 49K context · ~43 tok/s
- Tesla V100 32GB 32 GB Q4_K_M 49K context · ~43 tok/s
- L40S 48 GB Q8_0 53K context · ~24 tok/s
- RTX 6000 Ada 48 GB Q8_0 53K context · ~27 tok/s
- RTX A6000 48 GB Q8_0 53K context · ~21 tok/s
- RTX PRO 5000 Blackwell 48 GB Q8_0 53K context · ~37 tok/s
- RTX PRO 5000 Blackwell 72GB 72 GB Q8_0 128K context · ~37 tok/s
- RTX PRO 5500 Blackwell 84 GB Q8_0 128K context · ~39 tok/s
- RTX PRO 6000 Blackwell 96 GB Q8_0 128K context · ~50 tok/s
- DGX Spark (GB 10) 128 GB Q8_0 128K context · ~8 tok/s
AMD
- Radeon RX 7900 XTX 24 GB Q4_K_M 16K context · ~46 tok/s
- Radeon AI PRO R9700 32 GB Q4_K_M 49K context · ~31 tok/s
- Ryzen AI Max (Strix Halo) 32 GB Q4_K_M 49K context · ~12 tok/s
- Ryzen AI Max 400 (Gorgon Halo) 64 GB Q8_0 119K context · ~8 tok/s
Intel
- Arc Pro B65 32 GB Q4_K_M 49K context · ~29 tok/s
- Arc Pro B70 32 GB Q4_K_M 49K context · ~29 tok/s
Apple
- M2 24 GB Q4_K_M 16K context · ~5 tok/s
- M3 24 GB Q4_K_M 16K context · ~5 tok/s
- M4 24 GB Q4_K_M 16K context · ~6 tok/s
- M4 Pro 24 GB Q4_K_M 16K context · ~13 tok/s
- M5 24 GB Q4_K_M 16K context · ~7 tok/s
- M5 Pro 24 GB Q4_K_M 16K context · ~15 tok/s
- M6 24 GB Q4_K_M 16K context · ~8 tok/s
- M1 Max 32 GB Q4_K_M 49K context · ~19 tok/s
- M1 Pro 32 GB Q4_K_M 49K context · ~10 tok/s
- M2 Max 32 GB Q4_K_M 49K context · ~19 tok/s
- M2 Pro 32 GB Q4_K_M 49K context · ~10 tok/s
- M3 Max 36 GB Q4_K_M 65K context · ~14 tok/s
- M3 Pro 36 GB Q4_K_M 65K context · ~7 tok/s
- M4 Max 36 GB Q4_K_M 65K context · ~20 tok/s
- M5 Max 36 GB Q4_K_M 65K context · ~22 tok/s
- M1 Ultra 64 GB Q8_0 119K context · ~22 tok/s
- M2 Ultra 64 GB Q8_0 119K context · ~22 tok/s
- M3 Ultra 96 GB Q8_0 128K context · ~23 tok/s
- M5 Ultra 96 GB Q8_0 128K context · ~33 tok/s