Mac Studio M3 Ultra 256GBBudget pick
C228.2 tok/s decode
Can it run?
Runs well
Using Q4_K_M in Ollama
Fit status
Runs well
Decode
549.8 tok/s
TTFT
352 ms
Safe context
109K
Memory
14.0 GB / 96.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 549.8 tok/s | 512 ms | 212K |
| Chat | C | Runs well | 549.8 tok/s | 350 ms | 55K |
| Coding | C | Runs well | 549.8 tok/s | 352 ms | 109K |
| RAG | C | Runs well | 549.8 tok/s | 640 ms | 212K |
| Reasoning | C | Runs well | 549.8 tok/s | 416 ms | 109K |
How Qwen3.5 4B (4B params) fits at each quantization level on RTX PRO 6000 Blackwell Server Edition 96GB (96.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 1.6 GB | Low | D30 |
Q3_K_S | 3 | 2.0 GB | Low | D30 |
NVFP4 | 4 | 2.2 GB | Medium | D30 |
Q4_K_M | 4 | 2.4 GB | Medium | D30 |
Q5_K_M | 5 | 2.9 GB | High | D30 |
Q6_K | 6 | 3.3 GB | High | D30 |
Q8_0 | 8 | 4.3 GB | Very High | D30 |
F16Best for your GPU | 16 | 8.2 GB | Maximum | D31 |
Upgrade options