text model · Qwen3-Coder · macOS
Can I run Qwen3-Coder 480B-A35B Instruct on Apple M3 Ultra (256GB)?
No. Qwen3-Coder 480B-A35B Instruct needs ~276.2 GB even at Q4_K_M, but Apple M3 Ultra (256GB) only has ~192 GB usable.
Needs ~276.2 GB even at Q4_K_M, but only ~192 GB is usable.
That figure is at a 4k context and moves about ±15% as context length changes.
The gap is about 84.2 GB: Qwen3-Coder 480B-A35B Instruct needs roughly 276.2 GB at Q4_K_M and Apple M3 Ultra (256GB) leaves only about 192 GB usable for a model. No single tracked device has enough memory; Qwen3-Coder 480B-A35B Instruct needs a multi-GPU or high-memory rig.
- Q4_K_M needed
- ~276.2 GB
- Usable on device
- ~192 GB
- Device memory
- 256 GB
Which quant fits
- Parameters
- 480B (MoE, 35B active)
- Q4_K_M size
- 270.14 GB
- Q8_0 size
- 475.3 GB
- Context
- 256k
- Ollama tag
- qwen3-coder:480b
- Memory
- 256 GB unified
- Usable for weights
- ~192 GB
- Power draw
- ~270 W
- Best runtime
- MLX direct / Ollama (MLX backend)
What you can run instead
FAQ
Can Apple M3 Ultra (256GB) run Qwen3-Coder 480B-A35B Instruct?
No. Qwen3-Coder 480B-A35B Instruct needs ~276.2 GB even at Q4_K_M, but Apple M3 Ultra (256GB) only has ~192 GB usable.
How much memory does Qwen3-Coder 480B-A35B Instruct need?
Apple M3 Ultra (256GB) does not have enough memory. At Q4_K_M the weights are ~270.14 GB; with KV cache and runtime overhead, budget ~276.2 GB at a 4k context. It is a Mixture-of-Experts model (480B total / 35B active), so all experts must stay in memory; memory tracks total params, not active params.
What is the best tool to run Qwen3-Coder 480B-A35B Instruct on macOS?
LM Studio for a simple setup; mlx-lm for the most speed. vLLM is NOT a Mac tool, it is a CUDA/Linux serving engine. Unified memory is not a fixed VRAM slice; ~70% is usable for weights.
Embed this
A live badge for your model card or README, updated as the data is. [](https://localmodel.run/can-i-run/qwen3-coder-480b/apple-m3-ultra-256gb) Sources
Weights are measured from GGUF files; KV cache and overhead are computed, so totals can vary ~15% with context and runtime. Any tok/s is a bandwidth estimate. See methodology.