Mac Studio M3 Ultra 256GBBudget pick
C584.8 tok/s decode
Can it run?
Runs well
Using Q6_K in Ollama
Fit status
Runs well
Decode
2797.2 tok/s
TTFT
350 ms
Safe context
135K
Memory
15.2 GB / 128.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 2797.2 tok/s | 350 ms | 269K |
| Chat | C | Runs well | 2797.2 tok/s | 350 ms | 67K |
| Coding | C | Runs well | 2797.2 tok/s | 350 ms | 135K |
| RAG | C | Runs well | 2797.2 tok/s | 350 ms | 269K |
| Reasoning | C | Runs well | 2797.2 tok/s | 350 ms | 135K |
How SmolVLM 500M Instruct (0.5B params) fits at each quantization level on Gaudi 3 128GB (128.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 0.2 GB | Low | D29 |
Q3_K_S | 3 | 0.2 GB | Low | D29 |
NVFP4 | 4 |
huggingface-cli download hf-ggml-org--smolvlm-500m-instruct-ggufUpgrade options
0.3 GB |
| Medium |
| D29 |
Q4_K_M | 4 | 0.3 GB | Medium | D29 |
Q5_K_M | 5 | 0.4 GB | High | D30 |
Q6_K | 6 | 0.4 GB | High | D30 |
Q8_0 | 8 | 0.5 GB | Very High | D30 |
F16Best for your GPU | 16 | 1.0 GB | Maximum | D30 |