Mac Studio M3 Ultra 96GBBudget pick
C114.1 tok/s decode
Can it run?
Runs well
Using Q4_K_M in Ollama
Fit status
Runs well
Decode
161.3 tok/s
TTFT
1200 ms
Safe context
63K
Memory
12.1 GB / 48.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 161.3 tok/s | 1746 ms | 115K |
| Chat | C | Runs well | 161.3 tok/s | 655 ms | 33K |
| Coding | C | Runs well | 161.3 tok/s | 1200 ms | 63K |
| RAG | C | Runs well | 161.3 tok/s | 2182 ms | 115K |
| Reasoning | C | Runs well | 161.3 tok/s | 1418 ms | 63K |
How llava llama 3 8b v1 1 (8B params) fits at each quantization level on RTX 6000 Ada 48GB (48.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 3.1 GB | Low | D31 |
Q3_K_S | 3 | 3.9 GB | Low | D31 |
NVFP4 | 4 | 4.5 GB | Medium | D31 |
Q4_K_M | 4 | 4.9 GB | Medium | D32 |
Q5_K_M | 5 | 5.8 GB | High | D32 |
Q6_K | 6 | 6.6 GB | High | D32 |
Q8_0 | 8 | 8.6 GB | Very High | D33 |
F16Best for your GPU | 16 | 16.4 GB | Maximum | D37 |
Upgrade options