# Qwen3.5 122B-A10B: hardware requirements
> Qwen3.5 122B-A10B (Alibaba Qwen) is a 122B (10B active) mixture-of-experts model with a 256K-token context window, released under the Apache 2.0 license. It needs about 77GB of memory at Q4_K_M with 8K context.
Qwen3.5 122B-A10B is the mid-size mixture-of-experts model from Qwen's February 2026 Qwen3.5 launch, with 10B active parameters, image input, a 262K context, and switchable thinking. Later Qwen releases refreshed the 27B and 35B-A3B sizes, so this remains Qwen's main model for 96GB to 128GB machines. It needs about 77GB at Q4_K_M with 8K context, which suits Macs with 96GB or more, Strix Halo 128GB systems, and DGX Spark; a 24GB GPU needs 96GB or more of system RAM to hold the offloaded experts. Its hybrid attention keeps the KV cache small, so long contexts are cheap.

## Recommended hardware
- **Minimum GPU:** RTX PRO 5000 72GB (72GB), Q3_K_M, ~99 tokens/sec
- **Best value:** RTX PRO 5000 72GB (72GB), Q3_K_M, ~99 tokens/sec
- **Fastest single GPU:** RTX PRO 6000 (96GB), Q4_K_M, ~105 tokens/sec
- **Multi-GPU:** 4× RTX 3090 (96GB), Q4_K_M, ~39 tokens/sec
- **Multi-GPU:** 4× RTX 4090 (96GB), Q4_K_M, ~43 tokens/sec
- **Multi-GPU:** 3× RTX 5090 (96GB), Q4_K_M, ~66 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q4_K_M, ~18 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 45.8GB | 47.9GB | 48.5GB |
| Q3_K_M | 59.5GB | 62.1GB | 62.7GB |
| Q4_K_M | 74.0GB | 77.0GB | 77.6GB |
| Q5_K_M | 86.9GB | 90.3GB | 90.9GB |
| Q6_K | 100.6GB | 104.5GB | 105.1GB |
| Q8_0 | 129.6GB | 134.3GB | 134.9GB |
| BF16 | 244.0GB | 252.1GB | 252.7GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX PRO 6000 | 96GB | Q4_K_M | ~105 tok/s | ~$16,000 |
| RTX PRO 5000 72GB | 72GB | Q3_K_M | ~99 tok/s | ~$9,100 |
| M5 Max 128GB | 128GB | Q4_K_M | ~34 tok/s | ~$5,100 |
| M5 Ultra 512GB | 512GB | Q8_0 | ~33 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~33 tok/s | ~$9,499 |
| M4 Max 128GB | 128GB | Q4_K_M | ~31 tok/s | $3,499 MSRP |
| M3 Ultra 512GB | 512GB | Q8_0 | ~23 tok/s | $9,499 MSRP |
| Ryzen AI Max+ 395 | 128GB | Q4_K_M | ~18 tok/s | ~$3,300 |
| DGX Spark | 128GB | Q6_K | ~14 tok/s | ~$6,950 |
| Ryzen AI Max+ PRO 495 | 192GB | Q6_K | ~14 tok/s | ~$6,800 |

## How to run it
- Ollama: `ollama run qwen3.5:122b`
- llama.cpp: `llama-server -hf unsloth/Qwen3.5-122B-A10B-GGUF:Q4_K_M -c 8192 -ngl 99`
- Weights: https://huggingface.co/Qwen/Qwen3.5-122B-A10B

## Strengths
- Strong all-round model with image input and toggleable thinking
- 10B active parameters keep it fast on unified-memory machines
- Small KV cache thanks to hybrid linear attention
- Apache 2.0 license

## Weaknesses
- Needs about 77GB at Q4_K_M, beyond any single consumer GPU
- Smaller 2026 models such as Qwen3.8 27B now score about as well on benchmarks
- Hybrid attention needs recent runtime versions
Full page: https://llamabuilds.ai/models/qwen3-5-122b-a10b · Hardware guide: https://llamabuilds.ai/models/qwen3-5-122b-a10b/requirements

---
Source: https://llamabuilds.ai/models/qwen3-5-122b-a10b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.