Mac Studio M3 Ultra 96GBBudget pick
C77.4 tok/s decode
Can it run?
Runs well
Using Q4_K_M in Ollama
Fit status
Runs well
Decode
154.9 tok/s
TTFT
1250 ms
Safe context
38K
Memory
27.0 GB / 64.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 154.9 tok/s | 1818 ms | 75K |
| Chat | C | Runs well | 154.9 tok/s | 682 ms | 19K |
| Coding | C | Runs well | 154.9 tok/s | 1250 ms | 38K |
| RAG | C | Runs well | 154.9 tok/s | 2273 ms | 75K |
| Reasoning | C | Runs well | 154.9 tok/s | 1477 ms | 38K |
How Qwen3-Coder 30B A3B Instruct (30.5B params) fits at each quantization level on AMD Instinct MI210 64GB (64.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 11.9 GB | Low | D34 |
Q3_K_S | 3 | 14.9 GB | Low | D35 |
NVFP4 | 4 | 17.1 GB | Medium | D35 |
Q4_K_M | 4 | 18.6 GB | Medium | D36 |
Q5_K_M | 5 | 22.0 GB | High | D37 |
Q6_K | 6 | 25.0 GB | High | D38 |
Q8_0Best for your GPU | 8 | 32.6 GB | Very High | C40 |
F16 | 16 | 62.5 GB | Maximum | C45 |
ollama run qwen-3-coder-30b-a3bhuggingface-cli download qwen-3-coder-30b-a3bUpgrade options