Qwen 3.5 122B-A10B VRAM Requirements (2026)

Qwen 3.5 122B-A10B is the largest model in the Qwen 3.5 family: a 122B-parameter mixture-of-experts model that activates only about 10B parameters per token. That sparse activation is the whole story. It reasons at near-frontier level, decodes faster than a dense 122B, but still needs the full 72 GB of quantized weights in memory at Q4_K_M. On consumer hardware, 96 GB is the realistic entry point.

TL;DR: You need 72 GB for the Q4_K_M weights alone, plus context. Plan for 96 GB of VRAM or unified memory as the comfortable minimum. An RTX PRO 6000 96GB runs it at ~41 tok/s est.; a Mac Studio M5 Ultra 96GB at ~27 tok/s est. A 24GB gaming GPU cannot run it at any quant we recommend.

What "122B-A10B" Actually Means

The A10B suffix means roughly 10B parameters are active per token. That changes speed, not memory. Decode reads only the active expert weights per token, so generation feels like a much smaller model. Loading still requires memory for all 122B parameters.

This is why the model confuses sizing guides. It scores like a frontier model, runs at the speed of a 30B-class MoE, but occupies the disk and RAM of a 122B model. Treat it as a large model with a small-model speed profile.

VRAM Requirements by Quantization

ModelFit tracks the Q4_K_M build, which the engine sizes at 72 GB loaded. That is the quantization we recommend for this model class.

Context costs extra on top of weights. This is a hybrid linear-attention model, so its KV cache is far smaller than a dense model of the same size: about 24 KB per token at fp16, versus roughly 384 KB for a dense 100B+ model. A 32K context costs about 0.7 GB. Long-context work is affordable even on 96 GB.

Hardware That Actually Runs It

Speeds below are ModelFit engine estimates at Q4_K_M, labeled est. because we do not run benchmarks on this site.

The 5090 and 4090 rows deserve a warning: they technically load the model with CPU offload, but most of the weights sit in system RAM. Expect single-digit to low-teen tokens per second. Some sites list this model as "runs great" on a 24GB card. It does not.

What to Run Instead on Smaller Hardware

If you have 32-64 GB, the MoE design has smaller siblings that keep most of the quality:

• Qwen 3.6 35B-A3B (~22 GB at Q4) — same sparse design, 3B active, runs at ~39 tok/s est. on a 48GB Mac Mini M5 Pro and ~96 tok/s est. on an M5 Ultra.
• Qwen 3.5 35B-A3B (~20 GB at Q4) — previous generation of the same idea, slightly faster.
• GPT-OSS 120B (~65 GB native MXFP4) — the direct competitor at this tier, slightly lighter than Qwen 3.5 122B.

On 24GB GPUs, do not fight the 122B. A dense 27B like Qwen 3.8 27B at ~37 tok/s est. on an RTX 4090 will beat an offloaded 122B on both speed and usability.

FAQ

Can a 48GB GPU run Qwen 3.5 122B-A10B?

Not practically. Two RTX 4090s or a 48GB workstation card leave the model half-offloaded. The MoE architecture tolerates offload better than dense models, but you are still looking at speeds that negate the point of a 122B model. Use the 35B-A3B instead.

Is Q4_K_M good enough for this model?

Yes. At 122B total, Q4_K_M preserves nearly all of the benchmark quality. The quantization loss matters less at this scale than the context you can afford. Q8_0 is a marginal quality gain for 70% more memory.

How much context can I afford on 96 GB?

Plenty. After the 72 GB of weights and runtime overhead, you have roughly 15 GB of headroom. At 24 KB per token, that is enough for hundreds of thousands of tokens of context. Memory is not the constraint at 96 GB; the speed of the 10B active path is.

Does it run on a MacBook Pro?

Only on the 128GB M4 Max or M5 Max configs, and even then it is tight. The M5 Max 96GB Mac Studio is the cheapest Apple machine we would point at this model.