# Qwen3 30B-A3B (2507): hardware requirements
> Qwen3 30B-A3B (2507) (Alibaba Qwen) is a 30.5B (3.3B active) mixture-of-experts model with a 256K-token context window, released under the Apache 2.0 license. It needs about 21GB of memory at Q4_K_M with 8K context.
Qwen3 30B-A3B was the breakout local model of 2025: a 30.5B mixture-of-experts with only 3.3B active parameters, so it runs fast even with part of it offloaded to system RAM. This entry covers the July 2025 (2507) refresh, which split it into separate Instruct and Thinking checkpoints and raised native context to 262K. It has been superseded by Qwen3.5 35B-A3B and then Qwen3.6 35B-A3B, which add vision and stronger agentic coding at the same speed. At Q4_K_M it needs about 21GB with 8K context, so it fits a 24GB card, and a 12GB or 16GB GPU with 32GB or more of system RAM still runs it at usable speed.

## Recommended hardware
- **Minimum GPU:** RTX 4000 Ada (20GB), Q3_K_M, ~84 tokens/sec
- **Best value:** RTX 3090 (24GB), Q4_K_M, ~135 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), Q6_K, ~175 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q8_0, ~29 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 11.5GB | 13.3GB | 15.7GB |
| Q3_K_M | 15.0GB | 16.8GB | 19.3GB |
| Q4_K_M | 18.6GB | 20.6GB | 23.0GB |
| Q5_K_M | 21.9GB | 24.0GB | 26.4GB |
| Q6_K | 25.4GB | 27.5GB | 29.9GB |
| Q8_0 | 32.5GB | 34.9GB | 37.3GB |
| BF16 | 61.2GB | 64.4GB | 66.8GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | Q6_K | ~175 tok/s | ~$4,300 |
| RTX PRO 6000 | 96GB | Q8_0 | ~150 tok/s | ~$16,000 |
| RTX 4090 | 24GB | Q4_K_M | ~150 tok/s | ~$2,500 used |
| RTX 3090 | 24GB | Q4_K_M | ~135 tok/s | ~$1,050 used |
| RTX PRO 5000 72GB | 72GB | Q8_0 | ~125 tok/s | ~$9,100 |
| RX 7900 XTX | 24GB | Q4_K_M | ~110 tok/s | ~$1,050 used |
| RTX 6000 Ada | 48GB | Q8_0 | ~99 tok/s | ~$4,750 used |
| V100 32GB | 32GB | Q6_K | ~92 tok/s | ~$700 used |
| RTX 4000 Ada | 20GB | Q3_K_M | ~84 tok/s | $1,250 MSRP |
| M5 Ultra 512GB | 512GB | Q8_0 | ~80 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~80 tok/s | ~$9,499 |
| RTX A6000 | 48GB | Q8_0 | ~76 tok/s | ~$4,400 used |
| R9700 | 32GB | Q6_K | ~68 tok/s | ~$1,800 |
| MI50 32GB | 32GB | Q6_K | ~67 tok/s | ~$600 used |
| M3 Ultra 512GB | 512GB | Q8_0 | ~60 tok/s | $9,499 MSRP |
| Arc Pro B60 | 24GB | Q4_K_M | ~54 tok/s | ~$650 |
| M5 Max 128GB | 128GB | Q8_0 | ~54 tok/s | ~$5,100 |
| Arc Pro B70 | 32GB | Q6_K | ~54 tok/s | ~$1,300 |
| M4 Max 128GB | 128GB | Q8_0 | ~49 tok/s | $3,499 MSRP |
| Tesla P40 | 24GB | Q4_K_M | ~47 tok/s | ~$230 used |
| DGX Spark | 128GB | Q8_0 | ~31 tok/s | ~$6,950 |
| Ryzen AI Max+ PRO 495 | 192GB | Q8_0 | ~31 tok/s | ~$6,800 |
| Ryzen AI Max+ 395 | 128GB | Q8_0 | ~29 tok/s | ~$3,300 |

## How to run it
- Ollama: `ollama run qwen3:30b-a3b-instruct-2507-q4_K_M`
- llama.cpp: `llama-server -hf unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF:Q4_K_M -c 8192 -ngl 99`
- Weights: https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507

## Strengths
- Very fast generation: 3.3B active parameters per token
- Tolerates CPU offload well on 12GB to 16GB GPUs
- 262K native context with a modest 96 KiB per-token KV cache
- Separate Instruct and Thinking checkpoints, both Apache 2.0

## Weaknesses
- Superseded by Qwen3.6 35B-A3B for coding and agents
- Less world knowledge than dense models of the same total size
- Text only
Full page: https://llamabuilds.ai/models/qwen3-30b-a3b · Hardware guide: https://llamabuilds.ai/models/qwen3-30b-a3b/requirements

---
Source: https://llamabuilds.ai/models/qwen3-30b-a3b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.