Mac Studio M3 Ultra 256GBBudget pick
C101.4 tok/s decode
Can it run?
Runs well
Using Q4_K_M in Ollama
Fit status
Runs well
Decode
419.7 tok/s
TTFT
461 ms
Safe context
98K
Memory
20.9 GB / 128.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 396.4 tok/s | 710 ms | 184K |
| Chat | C | Runs well | 396.4 tok/s | 350 ms | 50K |
| Coding | C | Runs well | 419.7 tok/s | 461 ms | 98K |
| RAG | C | Runs well | 396.4 tok/s | 888 ms | 184K |
| Reasoning | C | Runs well | 396.4 tok/s | 577 ms | 98K |
How Qwen3.5 9B Uncensored HauhauCS Aggressive (9B params) fits at each quantization level on AMD Instinct MI250 128GB (128.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 3.5 GB | Low | D30 |
Q3_K_S | 3 | 4.4 GB | Low | D30 |
NVFP4 | 4 |
Upgrade options
5.0 GB |
| Medium |
| D30 |
Q4_K_M | 4 | 5.5 GB | Medium | D30 |
Q5_K_M | 5 | 6.5 GB | High | D31 |
Q6_K | 6 | 7.4 GB | High | D31 |
Q8_0 | 8 | 9.6 GB | Very High | D31 |
F16Best for your GPU | 16 | 18.5 GB | Maximum | D33 |