Qwen 3 235B A22B needs ~109.2 GB VRAM. Mac Studio M2 Ultra 128GB has 92.2 GB. With Q2_K quantization, expect ~10 tok/s.
Operating mode
Interactive favors responsiveness, while light API and scale-out lean harder on serving readiness. The fit stays the same, but the recommendation lens changes.
Current mode
Balanced
Balanced for general local use. Keeps the ranking neutral across personal and serving workflows.
Select quantization to explore
68.7 GB over capacity — needs offload or smaller quantization
Fit status
Too heavy
Decode
4.7 tok/s
TTFT
41549 ms
Safe context
4K
Memory
160.9 GB / 92.2 GB
Offload
40%
It fits through host-memory offload, and offload is the main reason performance drops.
CPU or host-memory offload is active
About 20% of the working set spills out of accelerator memory, which usually hurts latency and sustained decode throughput.
Very little memory headroom
You can run the model, but there is not much room left for longer context, bigger batches, extra apps, or future model updates.
Shared-memory contention still exists
The OS, browser, and inference runtime all compete for the same physical memory pool, so real-world headroom is less forgiving than raw capacity suggests.
Remove offload with more accelerator memory
Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Buy headroom, not only minimum fit
A slightly larger memory tier gives you safer context growth and makes the recommendation more future-proof.
Increase host RAM if you keep offloading
This setup may need roughly 14.3 GB of extra host RAM just for the offloaded portion, before OS and other tools.
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Chat | F | Too heavy | 4.7 tok/s | 22441 ms | 4K |
| Coding | F | Too heavy | 4.7 tok/s | 41549 ms | 4K |
| Agentic Coding | F | Too heavy | 4.6 tok/s | 61619 ms | 4K |
| Reasoning | F | Too heavy | 4.7 tok/s | 49103 ms | 4K |
| RAG | F | Too heavy | 4.6 tok/s | 77024 ms | 4K |
Inference speed
Estimated decode speed (tokens/sec) for Qwen 3 235B A22B at Q4_K_M across popular GPUs and Apple Silicon, including multi-GPU rigs, using the fastest local runtime per device. Fastest is Mac Studio M3 Ultra 256GB at ~11 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.
| GPU / Mac | Memory | Quant | Speed (tok/s) | Fits? |
|---|---|---|---|---|
Mac Studio M3 Ultra 256GB | 256 GB | Q4_K_M | 11.3 | Tight |
Mac Studio M2 Ultra 128GB | 128 GB | Q4_K_M | 4.7 | Too big |
Mac Studio M1 Ultra 128GB | 128 GB | Q4_K_M | 4.4 | Too big |
 RTX 5090 32GB | 32 GB | Q4_K_M | 3.7 | Too big |
2× RX 7900 XTX 24GB | 48 GB | Q4_K_M | 3.6 | Too big |
MacBook Pro M4 Max 128GB | 128 GB | Q4_K_M | 3.5 | Too big |
MacBook Pro M4 Max 64GB | 64 GB | Q4_K_M | 3.2 | Too big |
| 2× RTX 4090 24GB | 48 GB | Q4_K_M | 2.5 | Too big |
| RTX 4090 24GB | 24 GB | Q4_K_M | 2.3 | Too big |
MacBook Pro M3 Max 64GB | 64 GB | Q4_K_M | 2.2 | Too big |
RX 7900 XTX 24GB | 24 GB | Q4_K_M | 2.1 | Too big |
| 2× RTX 3090 24GB | 48 GB | Q4_K_M | 2.1 | Too big |
| RTX 4080 Super 16GB | 16 GB | Q4_K_M | 2.0 | Too big |
| RTX 3090 24GB | 24 GB | Q4_K_M | 2.0 | Too big |
| RTX 4070 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 3060 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 4060 8GB | 8 GB | Q4_K_M | 2.0 | Too big |
MacBook Pro M1 Max 64GB | 64 GB | Q4_K_M | 2.0 | Too big |
MacBook Pro M4 Pro 48GB | 48 GB | Q4_K_M | 2.0 | Too big |
| 4× RTX 3060 12GB | 48 GB | Q4_K_M | 2.0 | Too big |
Estimates for single-stream decoding at Q4_K_M; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.
How Qwen 3 235B A22B (235B params) fits at each quantization level on Mac Studio M2 Ultra 128GB (92.2 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 91.7 GB | Low | F0 |
Q3_K_S | 3 | 115.2 GB | Low | F0 |
NVFP4 | 4 | 131.6 GB | Medium | F0 |
Q4_K_M | 4 | 143.4 GB | Medium | F0 |
Q5_K_M | 5 | 169.2 GB | High | F0 |
Q6_K | 6 | 192.7 GB | High | F0 |
Q8_0 | 8 | 251.5 GB | Very High | F0 |
F16 | 16 | 481.7 GB | Maximum | F0 |
Copy-paste commands to run Qwen 3 235B A22B on your machine.
Run
lms load Qwen3-235B-A22B-Instruct-2507 && lms server startOpciones de mejora
Hace que el modelo quepa en el acelerador en lugar de seguir fuera de alcance.
Elimina el offload a memoria del sistema, que suele ser la mayor mejora individual en latencia y throughput.
~$6,999 MSRP
Hace que el modelo quepa en el acelerador en lugar de seguir fuera de alcance.
Elimina el offload a memoria del sistema, que suele ser la mayor mejora individual en latencia y throughput.
~$8,000 MSRP
Yes, Mac Studio M2 Ultra 128GB can run Qwen 3 235B A22B at Q2_K quantization (Very compromised (needs ~14.3 GB host RAM)). The recommended Q4_K_M requires 160.9 GB which exceeds available memory, but at Q2_K it needs only 109.2 GB. Expected decode speed: 9.7 tok/s.
Qwen 3 235B A22B (235B parameters) requires approximately 160.9 GB at Q4_K_M quantization. On Mac Studio M2 Ultra 128GB, it fits at Q2_K using 109.2 GB.
The recommended quantization is Q4_K_M, but on Mac Studio M2 Ultra 128GB the best fitting quantization is Q2_K, which uses 109.2 GB.
On Mac Studio M2 Ultra 128GB, Qwen 3 235B A22B achieves approximately 9.7 tokens per second decode speed with a time-to-first-token of 19939ms using Q2_K quantization.
For coding workloads, Qwen 3 235B A22B on Mac Studio M2 Ultra 128GB receives a F grade with 4.7 tok/s and 4K context.
On Mac Studio M2 Ultra 128GB, Qwen 3 235B A22B can safely use up to 4K tokens of context at Q2_K quantization. The model's official context limit is 131K, but available memory constrains the safe maximum.
Remove offload with more accelerator memory. Prioritize a GPU or unified-memory tier that fits the whole model natively. Removing offload usually helps more than small compute gains.
Not always. Mac Studio M2 Ultra 128GB can often fit larger models thanks to unified memory, but a discrete GPU with dedicated high-bandwidth VRAM may still decode faster once the model fits. For this combination, the important distinction is capacity versus sustained throughput.
Paste this snippet into any page to show a live fit card.
<iframe src="https://willitrunai.com/embed/qwen-3-235b-a22b-on-m2-ultra-128gb" width="400" height="180" frameborder="0" style="border:none;border-radius:12px;overflow:hidden;" title="Will It Run AI — fit result"></iframe>
Preview: