Mac Studio M3 Ultra 96GBBudget pick
C130.4 tok/s decode
Can it run?
Runs well
Using Q4_K_M in Ollama
Fit status
Runs well
Decode
161.9 tok/s
TTFT
1196 ms
Safe context
68K
Memory
11.4 GB / 48.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 184.4 tok/s | 1527 ms | 123K |
| Chat | C | Runs well | 184.4 tok/s | 573 ms | 35K |
| Coding | C | Runs well | 161.9 tok/s | 1196 ms | 68K |
| RAG | C | Runs well | 184.4 tok/s | 1909 ms | 123K |
| Reasoning | C | Runs well | 184.4 tok/s | 1241 ms | 68K |
How Mistral 7B Instruct v0.2 (7B params) fits at each quantization level on RTX 6000 Ada 48GB (48.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 2.7 GB | Low | D31 |
Q3_K_S | 3 | 3.4 GB | Low | D31 |
NVFP4 | 4 |
Upgrade options
3.9 GB |
| Medium |
| D31 |
Q4_K_M | 4 | 4.3 GB | Medium | D31 |
Q5_K_M | 5 | 5.0 GB | High | D32 |
Q6_K | 6 | 5.7 GB | High | D32 |
Q8_0 | 8 | 7.5 GB | Very High | D33 |
F16Best for your GPU | 16 | 14.3 GB | Maximum | D36 |