RTX A6000 48GBBudget pickC43.5 tok/s decode
~$4,650 MSRP
Can it run?
Runs well
Using Q4_K_M in vLLM
Fit status
Runs well
Decode
17.9 tok/s
TTFT
10825 ms
Safe context
28K
Memory
26.2 GB / 46.1 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 17.9 tok/s | 15746 ms | 50K |
| Chat | C | Runs well | 17.9 tok/s | 5905 ms | 15K |
| Coding | C | Runs well | 17.9 tok/s | 10825 ms | 28K |
| RAG | C | Runs well | 17.9 tok/s | 19683 ms | 50K |
| Reasoning | C | Runs well | 19.4 tok/s | 11810 ms | 28K |
How Codestral 2 25.08 (22B params) fits at each quantization level on MacBook Pro M3 Max 64GB (46.1 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.6 GB | Low | D33 |
Q3_K_S | 3 | 10.8 GB | Low | D34 |
NVFP4 | 4 |
huggingface-cli download codestral-2-25.08Upgrade options
RTX A6000 48GBBudget pick~$4,650 MSRP
NVIDIA L40 48GBBest value~$5,500 MSRP
RTX PRO 5000 Blackwell 48GBBiggest leap
12.3 GB |
| Medium |
| D35 |
Q4_K_M | 4 | 13.4 GB | Medium | D36 |
Q5_K_M | 5 | 15.8 GB | High | D37 |
Q6_K | 6 | 18.0 GB | High | D38 |
Q8_0Best for your GPU | 8 | 23.5 GB | Very High | C40 |
F16 | 16 | 45.1 GB | Maximum | C44 |