# Qwen3-8B

> OnPrem World · https://onpremworld.com/en/models/qwen3-8b · Last updated: 2026-10-16 · CC BY 4.0

Qwen's 8B model with a thinking mode you can switch off, good at Chinese; it fits a 12 GB card at 8-bit. Native context is 32K (128K with YaRN).

|  |  |
|---|---|
| Developer | Alibaba Qwen |
| Parameters | 8.2B |
| Native context | 32,768 tokens |
| Licence | Apache-2.0 |
| Commercial use | Yes |
| KV cache | All 36 layers (standard attention) |
| Released | 2025-04 |

```
ollama run qwen3:8b
```

## Memory needed

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

| Quantization | 4K | 8K | 32K |
|---|---|---|---|
| Q4_K_M | 5.9 GiB | 6.5 GiB | 9.9 GiB |
| Q5_K_M | 6.8 GiB | 7.3 GiB | 10.7 GiB |
| Q6_K | 7.7 GiB | 8.2 GiB | 11.6 GiB |
| Q8_0 | 9.6 GiB | 10.1 GiB | 13.5 GiB |
| FP16 | 17.1 GiB | 17.7 GiB | 21.0 GiB |

## Which hardware runs it (8K)

| Hardware | Memory | Best quant | tok/s (est.) |
|---|---|---|---|
| DGX Station (GB300) | 252 GB | FP16 | 189 |
| GeForce RTX 5080 | 16 GB | Q8_0 | 67 |
| GeForce RTX 5090 | 32 GB | FP16 | 66 |
| RTX PRO 6000 Blackwell Workstation Edition | 96 GB | FP16 | 66 |
| RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96 GB | FP16 | 66 |
| GeForce RTX 5070 Ti | 16 GB | Q8_0 | 63 |
| GeForce RTX 4080 SUPER | 16 GB | Q8_0 | 53 |
| RTX PRO 5000 72GB Blackwell | 72 GB | FP16 | 51 |
| RTX PRO 5000 48GB Blackwell | 48 GB | FP16 | 51 |
| GeForce RTX 5070 | 12 GB | Q8_0 | 49 |
| Radeon RX 9070 XT | 16 GB | Q8_0 | 47 |
| M5 Ultra | 512 GB | FP16 | 43 |
| GeForce RTX 4090 | 24 GB | FP16 | 40 |
| Radeon RX 7900 XTX | 24 GB | FP16 | 38 |
| RTX 6000 Ada Generation | 48 GB | FP16 | 38 |
| GeForce RTX 3090 | 24 GB | FP16 | 37 |
| RTX PRO 4500 Blackwell | 32 GB | FP16 | 36 |
| Arc B580 | 12 GB | Q8_0 | 34 |
| GeForce RTX 5060 Ti 16GB | 16 GB | Q8_0 | 34 |
| Radeon RX 7900 XT | 20 GB | FP16 | 32 |
| RTX A6000 | 48 GB | FP16 | 31 |
| M3 Ultra | 512 GB | FP16 | 30 |
| GeForce RTX 3060 (12 GB) | 12 GB | Q8_0 | 27 |
| RTX PRO 4000 Blackwell | 24 GB | FP16 | 27 |
| Radeon AI PRO R9700 | 32 GB | FP16 | 26 |
| Arc Pro B70 | 32 GB | FP16 | 25 |
| Radeon RX 9060 XT 16GB | 16 GB | Q8_0 | 24 |
| M5 Max (18-core CPU / 40-core GPU) | 128 GB | FP16 | 23 |
| GeForce RTX 4060 Ti 16GB | 16 GB | Q8_0 | 22 |
| M4 Max (16-core CPU / 40-core GPU) | 128 GB | FP16 | 21 |
| Arc Pro B60 | 24 GB | FP16 | 19 |
| M5 Max (18-core CPU / 32-core GPU) | 36 GB | FP16 | 18 |
| M4 Max (14-core CPU / 32-core GPU) | 36 GB | FP16 | 16 |
| M5 Pro | 64 GB | FP16 | 12 |
| M6 (16GB configuration, 153 GB/s) | 16 GB | Q8_0 | 11 |
| RTX Spark (N1X) laptops & compact desktops | 128 GB | FP16 | 11 |
| M4 Pro | 64 GB | FP16 | 11 |
| DGX Spark | 128 GB | FP16 | 9.8 |
| Ryzen AI Max+ PRO 495 (Gorgon Halo) systems | 192 GB | FP16 | 9.8 |
| Ryzen AI Max+ 395 (Strix Halo) systems | 128 GB | FP16 | 9.2 |
| M6 (24GB/32GB configurations, 170 GB/s) | 32 GB | FP16 | 6.6 |
| M5 | 32 GB | FP16 | 6.0 |

## Frequently asked questions

### How much VRAM does Qwen3-8B need?

With an 8K context, about 6.5 GiB at Q4_K_M and 10.1 GiB at Q8_0 (weights + KV cache + runtime overhead). Longer contexts need more memory for the KV cache.

### Can Qwen3-8B run on an GeForce RTX 4090?

Yes. Qwen3-8B fits on the GeForce RTX 4090 (24 GB) at FP16 with an 8K context, using about 17.7 GiB, at an estimated 40 tokens/s.

### Can I run Qwen3-8B on a Mac?

Yes. Qwen3-8B fits on the M3 Ultra (96 GB) at FP16 with an 8K context, using about 17.7 GiB, at an estimated 30 tokens/s.

### Can Qwen3-8B be used commercially?

Qwen3-8B is released under Apache-2.0. The licence allows commercial use (read its conditions).

## Sources

- https://github.com/QwenLM/Qwen3/blob/main/README.md
- https://huggingface.co/Qwen/Qwen3-8B/blob/main/LICENSE
- https://huggingface.co/Qwen/Qwen3-8B
- https://huggingface.co/Qwen/Qwen3-8B/resolve/main/config.json
- https://huggingface.co/api/models/Qwen/Qwen3-8B?expand[]=safetensors
- https://ollama.com/library/qwen3/tags