Qwen3.6-35B-A3B (MoE) on Apple Mac mini M4 (24GB Unified): VRAM Requirement & Speed Estimate
A single Mac mini M4 24G (18.7 GB usable VRAM) cannot fully fit Qwen3.6-35B-A3B (MoE) at Q4_K_M (23.96 GB required). Consider multi-GPU tensor parallelism, a smaller parameter variant, or cloud GPU / API deployment.
Physical 24 GB · Load 128%
Exceeds usable VRAM by 5.24 GB
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3-6-35b-a3bInsufficient Local VRAM? Ways to Run Qwen3.6-35B-A3B (MoE)
Rent an hourly high-VRAM cloud GPU pod, or check the 12-month Buy vs Cloud vs API break-even curve.
📊 All Quantization Levels for Qwen3.6-35B-A3B (MoE) on Mac mini M4 24G
| Quantization | Total VRAM (16K) | Verdict | Est. Speed | Max Context |
|---|---|---|---|---|
| FP16 / BF16 | 75.14 GB | Out of VRAM (OOM) | — | — |
| FP8 / Q8_0 | 40.72 GB | Out of VRAM (OOM) | — | — |
| GGUF Q6_K | 31.81 GB | Out of VRAM (OOM) | — | — |
| GGUF Q5_K_M | 27.82 GB | Out of VRAM (OOM) | — | — |
| Q4_K_M / AWQ | 23.96 GB | Out of VRAM (OOM) | — | — |
| GGUF IQ4_XS | 21.21 GB | Out of VRAM (OOM) | — | — |
| GGUF Q3_K_M | 19.65 GB | Out of VRAM (OOM) | — | — |
| GGUF Q2_K | 16.16 GB | Tight Fit (Watch Context) | ~52.7 tok/s | 138K |
💡 Recommended Alternative GPUs & Macs for Qwen3.6-35B-A3B (MoE)
Frequently Asked Questions
Can Apple Mac mini M4 (24GB Unified) run Qwen3.6-35B-A3B (MoE) locally?
A single Mac mini M4 24G (18.7 GB usable VRAM) cannot fully fit Qwen3.6-35B-A3B (MoE) at Q4_K_M (23.96 GB required). Consider multi-GPU tensor parallelism, a smaller parameter variant, or cloud GPU / API deployment.
What is the best quantization for Qwen3.6-35B-A3B (MoE) on Mac mini M4 24G?
The highest-quality quantization that fits cleanly on Mac mini M4 24G is GGUF Q2_K (Extreme Compression), using 16.16 GB VRAM at ~52.7 tokens/s with up to 138K context.
How much VRAM does Qwen3.6-35B-A3B (MoE) need in full FP16 precision?
In unquantized FP16/BF16 precision (16K context), Qwen3.6-35B-A3B (MoE) requires 75.14 GB VRAM (72 GB for weights alone). Because it uses a Mixture-of-Experts (MoE) architecture, all 36B parameters must sit in memory, while only 3B active parameters are read per token.