Hardware guide · October 2026
What GPU do you need to run Qwen3.5 122B-A10B?
Qwen3.5 122B-A10B is a 122B (10B active) mixture-of-experts model. Here is exactly how much memory it needs and which GPUs and Macs run it well.
The short answer
To run Qwen3.5 122B-A10B at Q4_K_M with 8K tokens of context you need about 77GB of VRAM (74.0GB of weights plus KV cache and overhead). The cheapest way in is a 72GB card like the RTX PRO 5000 72GB (~$9,100). On a Mac or other unified-memory system, the Ryzen AI Max+ 395 runs it at ~18 tokens/sec.
- VRAM at Q4_K_M
- ~77GB
- GPUs that run it
- 10
- Parameters
- 122B
- CPU only (est.)
- ~7.2 tok/s
Our recommendations
The best GPUs for Qwen3.5 122B-A10B
Minimum
RTX PRO 5000 72GB72GB · ~$9,100
- Quant
- Q3_K_M
- Speed (est.)
- ~99 tok/s
The smallest card that runs it at Q3_K_M with 8K context.
Fastest
RTX PRO 600096GB · ~$16,000
- Quant
- Q4_K_M
- Speed (est.)
- ~105 tok/s
The fastest single GPU in our database for this model.
Multi-GPU
4× RTX 309096GB total · ~$4,200 used
- Quant
- Q4_K_M
- Speed (est.)
- ~39 tok/s
Split across cards with llama.cpp, vLLM, or ExLlama. Needs a board and PSU that can feed every card.
Faster multi-GPU
4× RTX 409096GB total · ~$10,000 used
- Quant
- Q4_K_M
- Speed (est.)
- ~43 tok/s
Split across cards with llama.cpp, vLLM, or ExLlama. Needs a board and PSU that can feed every card.
Mac / unified memory
Ryzen AI Max+ 395128GB unified · ~$3,300
- Quant
- Q4_K_M
- Speed (est.)
- ~18 tok/s
Lots of memory per dollar and silent, but slower than a GPU with the same capacity.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Memory
Qwen3.5 122B-A10B VRAM requirements
Memory needed at each quantization level. Q4_K_M is the best balance of size and quality for most people; move up to Q5 or Q6 if you have the room.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| Q2_K | Rough | 46 GB | 48 GB | 49 GB |
| Q3_K_M | Usable | 59 GB | 62 GB | 63 GB |
| Q4_K_M | RecommendedDefault | 74 GB | 77 GB | 78 GB |
| Q5_K_M | Very good | 87 GB | 90 GB | 91 GB |
| Q6_K | Near-lossless | 101 GB | 104 GB | 105 GB |
| Q8_0 | Practically lossless | 130 GB | 134 GB | 135 GB |
| BF16 | Full precision | 244 GB | 252 GB | 253 GB |
Context
How context length changes the memory you need
Every token in the context window is stored in the KV cache. At Q4_K_M, with a 16-bit KV cache, Qwen3.5 122B-A10B needs:
| Context | KV cache | Total memory | With 8-bit KV |
|---|---|---|---|
| 4K tokens | 0.1 GB | 76.9 GB | 76.8 GB |
| 8K tokens | 0.2 GB | 77.0 GB | 76.9 GB |
| 16K tokens | 0.4 GB | 77.2 GB | 77.0 GB |
| 32K tokens | 0.8 GB | 77.6 GB | 77.2 GB |
| 64K tokens | 1.6 GB | 78.4 GB | 77.6 GB |
| 128K tokens | 3.2 GB | 80.0 GB | 78.4 GB |
| 256K tokens | 6.4 GB | 83.2 GB | 80.0 GB |
Hybrid layout: three of every four layers use Gated DeltaNet linear attention with a fixed-size state. Assuming 48 layers with 12 full-attention layers of 2 KV heads x 256 dims: 2 x 12 x 2 x 256 x 2 bytes = 24,576 bytes per token, about 6.4GB at the full 262K.
Compatibility
Every GPU that runs Qwen3.5 122B-A10B
Each GPU at the highest-quality quantization that fits with 8K context, fastest first. Speeds are single-user estimates.
| GPU | Memory | Quant | Speed | Fit | Price |
|---|---|---|---|---|---|
| RTX PRO 6000 | 96GB | Q4_K_M | ~105 tok/s | Runs well | ~$16,000 |
| RTX PRO 5000 72GB | 72GB | Q3_K_M | ~99 tok/s | Runs well | ~$9,100 |
| M5 Max 128GB | 128GB unified | Q4_K_M | ~34 tok/s | Runs well | ~$5,100 |
| M5 Ultra 512GB | 512GB unified | Q8_0 | ~33 tok/s | Runs well | — |
| M5 Ultra 256GB | 256GB unified | Q8_0 | ~33 tok/s | Runs well | ~$9,499 |
| M4 Max 128GB | 128GB unified | Q4_K_M | ~31 tok/s | Runs well | $3,499 |
| M3 Ultra 512GB | 512GB unified | Q8_0 | ~23 tok/s | Runs well | $9,499 |
| Ryzen AI Max+ 395 | 128GB unified | Q4_K_M | ~18 tok/s | Runs well | ~$3,300 |
| Ryzen AI Max+ PRO 495 | 192GB unified | Q6_K | ~14 tok/s | Runs well | ~$6,800 |
| DGX Spark | 128GB unified | Q6_K | ~14 tok/s | Runs well | ~$6,950 |
With CPU offload (partial offload, assuming 64GB of system RAM): RTX 6000 Ada (~16 tok/s), RTX A6000 (~15 tok/s), RTX 5090 (~11 tok/s), V100 32GB (~11 tok/s), R9700 (~10 tok/s), MI50 32GB (~10 tok/s), Arc Pro B70 (~10 tok/s).
Tips
Getting the most out of Qwen3.5 122B-A10B
- Start at Q4_K_M. It keeps nearly all of the quality at about a third of the 16-bit size. Step up to Q5_K_M or Q6_K only if it still fits with the context you need.
- Budget memory for context. The KV cache grows with every token. Quantizing it to 8-bit roughly halves that cost with little quality loss.
- Offload experts, not layers. For mixture-of-experts models, llama.cpp's
--n-cpu-moe(or-ottensor overrides) keeps attention on the GPU and moves expert weights to system RAM. That keeps speed usable when the model does not fit in VRAM. - Pick the right runtime. Ollama and LM Studio are easiest. llama.cpp gives the most control. vLLM or SGLang are best when several people or agents share one machine.
Want to try different quantizations, context lengths, or GPU counts? Use our VRAM calculator, or read the Qwen3.5 122B-A10B overview for setup commands and benchmarks.
Frequently asked questions
How much VRAM do I need to run Qwen3.5 122B-A10B?
About 77GB at Q4_K_M with an 8K context window. At Q8_0 (near-lossless) it needs about 134GB, and at Q3_K_M about 62GB with some quality loss. Longer contexts add 0.03GB per 1K tokens with a 16-bit KV cache.
Can I run Qwen3.5 122B-A10B on a 24GB GPU?
No. Qwen3.5 122B-A10B needs about 77GB at Q4_K_M, which does not fit on a 24GB card like the RTX 3090 or RTX 4090, even with CPU offload on a 64GB system.
Can I run Qwen3.5 122B-A10B on a Mac?
Yes. With 128GB of unified memory or more (for example the Ryzen AI Max+ 395), macOS can give the GPU enough memory to run it at Q4_K_M, at about ~18 tokens/sec. MLX builds are usually fastest on Apple silicon.
Can I run Qwen3.5 122B-A10B on CPU only, without a GPU?
Yes, if you have at least 96GB of system RAM. On a desktop with dual-channel DDR5, expect roughly 7.2 tokens/sec, which is better than you might expect because only the active experts are read for each token. A GPU is strongly recommended for interactive use.