# DeepSeek-R1-Distill-Qwen-32B

> OnPrem World · https://onpremworld.com/en/models/deepseek-r1-distill-qwen-32b · Last updated: 2026-10-14 · CC BY 4.0

DeepSeek-R1's reasoning distilled into Qwen2.5-32B: a dense 32B that fits a 24 GB card at 4-bit.

|  |  |
|---|---|
| Developer | DeepSeek |
| Parameters | 32.76B |
| Native context | 131,072 tokens |
| Licence | MIT |
| Commercial use | Yes |
| KV cache | All 64 layers (standard attention) |
| Released | 2025-01 |

```
ollama run deepseek-r1:32b
```

## Memory needed

Weights + KV cache + runtime overhead, in GiB. Lower quantization = less memory, slightly lower quality.

| Quantization | 4K | 8K | 32K |
|---|---|---|---|
| Q4_K_M | 20.9 GiB | 21.9 GiB | 27.9 GiB |
| Q5_K_M | 24.3 GiB | 25.3 GiB | 31.3 GiB |
| Q6_K | 27.9 GiB | 28.9 GiB | 34.9 GiB |
| Q8_0 | 35.5 GiB | 36.5 GiB | 42.5 GiB |
| FP16 | 65.6 GiB | 66.6 GiB | 72.6 GiB |

## Which hardware runs it (8K)

| Hardware | Memory | Best quant | tok/s (est.) |
|---|---|---|---|
| GeForce RTX 5090 | 32 GB | Q6_K | 46 |
| GeForce RTX 4090 | 24 GB | Q4_K_M | 36 |
| Radeon RX 7900 XTX | 24 GB | Q4_K_M | 34 |
| GeForce RTX 3090 | 24 GB | Q4_K_M | 33 |
| RTX 6000 Ada Generation | 48 GB | Q8_0 | 19 |
| RTX PRO 6000 Blackwell Workstation Edition | 96 GB | FP16 | 19 |
| Radeon AI PRO R9700 | 32 GB | Q6_K | 17 |
| Arc Pro B60 | 24 GB | Q4_K_M | 16 |
| Arc Pro B70 | 32 GB | Q6_K | 16 |
| RTX A6000 | 48 GB | Q8_0 | 15 |
| RTX PRO 5000 72GB Blackwell | 72 GB | FP16 | 14 |
| M5 Max (18-core CPU / 32-core GPU) | 36 GB | Q5_K_M | 13 |
| M5 Ultra | 512 GB | FP16 | 12 |
| M4 Max (14-core CPU / 32-core GPU) | 36 GB | Q5_K_M | 11 |
| M3 Ultra | 512 GB | FP16 | 8.1 |
| M5 Max (18-core CPU / 40-core GPU) | 128 GB | FP16 | 6.1 |
| M5 Pro | 64 GB | Q8_0 | 5.7 |
| M6 (24GB/32GB configurations, 170 GB/s) | 32 GB | Q4_K_M | 5.6 |
| M4 Max (16-core CPU / 40-core GPU) | 128 GB | FP16 | 5.4 |
| M4 Pro | 64 GB | Q8_0 | 5.1 |
| M5 | 32 GB | Q4_K_M | 5.0 |
| RTX Spark (N1X) laptops & compact desktops | 128 GB | FP16 | 2.7 |
| DGX Spark | 128 GB | FP16 | 2.5 |
| Ryzen AI Max+ PRO 495 (Gorgon Halo) systems | 192 GB | FP16 | 2.5 |
| Ryzen AI Max+ 395 (Strix Halo) systems | 128 GB | FP16 | 2.3 |
| GeForce RTX 3060 (12 GB) | 12 GB | Doesn't fit | – |
| GeForce RTX 4060 Ti 16GB | 16 GB | Doesn't fit | – |
| GeForce RTX 4080 SUPER | 16 GB | Doesn't fit | – |
| GeForce RTX 5060 Ti 16GB | 16 GB | Doesn't fit | – |
| GeForce RTX 5070 | 12 GB | Doesn't fit | – |
| GeForce RTX 5070 Ti | 16 GB | Doesn't fit | – |
| GeForce RTX 5080 | 16 GB | Doesn't fit | – |
| Radeon RX 7900 XT | 20 GB | Doesn't fit | – |
| Radeon RX 9060 XT 16GB | 16 GB | Doesn't fit | – |
| Radeon RX 9070 XT | 16 GB | Doesn't fit | – |
| Arc B580 | 12 GB | Doesn't fit | – |
| M6 (16GB configuration, 153 GB/s) | 16 GB | Doesn't fit | – |

## Frequently asked questions

### How much VRAM does DeepSeek-R1-Distill-Qwen-32B need?

With an 8K context, about 21.9 GiB at Q4_K_M and 36.5 GiB at Q8_0 (weights + KV cache + runtime overhead). Longer contexts need more memory for the KV cache.

### Can DeepSeek-R1-Distill-Qwen-32B run on an GeForce RTX 4090?

Yes. DeepSeek-R1-Distill-Qwen-32B fits on the GeForce RTX 4090 (24 GB) at Q4_K_M with an 8K context, using about 21.9 GiB, at an estimated 36 tokens/s.

### Can I run DeepSeek-R1-Distill-Qwen-32B on a Mac?

Yes. DeepSeek-R1-Distill-Qwen-32B fits on the M3 Ultra (96 GB) at FP16 with an 8K context, using about 66.6 GiB, at an estimated 8.1 tokens/s.

### Can DeepSeek-R1-Distill-Qwen-32B be used commercially?

DeepSeek-R1-Distill-Qwen-32B is released under MIT. The licence allows commercial use (read its conditions).

## Sources

- https://api-docs.deepseek.com/news/news250120
- https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
- https://huggingface.co/api/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B?expand[]=safetensors
- https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/resolve/main/config.json
- https://ollama.com/library/deepseek-r1/tags