Models · chat and writing, coding, reasoning and maths
Qwen3 235B-A22B (MoE): VRAM requirements and GPUs that run it
Frontier-class open model. Mac Studio 256 GB territory. At Q4 with 8K context it needs about 155 GB, so you want a GPU with at least 192 GB of VRAM or a Mac with 192 GB of unified memory.
Memory needed by quantization
| Quant | Weights | Total at 8K | Total at 32K |
|---|---|---|---|
| Q4_K_M | 153 GB | 155 GB | 164 GB |
| Q5_K_M | 184 GB | 186 GB | 195 GB |
| Q8_0 | 274 GB | 277 GB | 286 GB |
| FP16 | 517 GB | 520 GB | 529 GB |
Totals include about 10% runtime overhead and the KV cache. Longer conversations cost real memory: going from 8K to 32K context adds about 8.8 GB at any quantization. It is a mixture-of-experts model: 235B parameters in total but only about 22B are used for each token, so it runs much faster than a dense model of the same size and stays usable even when part of it sits in system RAM. Most people should download Q4_K_M; Q5 or Q8 only make sense if you have spare memory. Read which quantization to pick.
Which GPUs can run Qwen3 235B-A22B (MoE)?
No single consumer GPU we track runs it fully in VRAM at 8K context.
Running it on a Mac
Apple Silicon Macs can give a model about 72% of their unified memory. Qwen3 235B-A22B (MoE) fits from 192 GB (Apple M2 Ultra, about 19 tokens per second). Compare the options in our Mac vs NVIDIA guide.
Compare Qwen3 235B-A22B (MoE)
Questions
How much VRAM does Qwen3 235B-A22B (MoE) need?
About 155 GB at Q4_K_M with 8K context, 277 GB at Q8_0, and 520 GB at full FP16. In practice that means a card with at least 192 GB of VRAM for Q4.
Can Qwen3 235B-A22B (MoE) run on an 8 GB GPU?
No. It needs about 155 GB at Q4, well beyond 8 GB even with offload.
Can I run Qwen3 235B-A22B (MoE) on a Mac?
Yes. The smallest configuration that runs it fully is Apple M2 Ultra with 192 GB of unified memory, at about 19 tokens per second. More memory bandwidth (Pro, Max, Ultra chips) means faster replies.
Estimates based on model size, architecture and memory bandwidth. Real numbers vary with software, drivers and context length.