RX 7900 XTX 24GBBudget pick
C51.5 tok/s decode
~$999 MSRP
Can it run?
Runs with offload
Using Q4_K_M in vLLM
Fit status
Runs with offload
Decode
19.4 tok/s
TTFT
9993 ms
Safe context
16K
Memory
22.7 GB / 23.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | D | Very compromised (needs ~1.6 GB host RAM) | 16.7 tok/s | 16892 ms | 28K |
| Chat | C | Tight fit | 17.3 tok/s | 6108 ms | 9K |
| Coding | C | Runs with offload | 19.4 tok/s | 9993 ms | 16K |
| RAG | D | Very compromised (needs ~1.6 GB host RAM) | 16.7 tok/s | 21115 ms | 28K |
| Reasoning | C | Runs with offload | 17.3 tok/s | 13235 ms |
How Codestral 2 25.08 (22B params) fits at each quantization level on MacBook Pro M2 Max 32GB (23.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.6 GB | Low | D37 |
Q3_K_S | 3 | 10.8 GB | Low | D39 |
NVFP4 | 4 |
huggingface-cli download codestral-2-25.08Upgrade options
~$999 MSRP
RTX 3090 24GBBest value~$1,499 MSRP
RTX 5090 32GBBiggest leap~$1,999 MSRP
| 16K |
12.3 GB |
| Medium |
| C41 |
Q4_K_M | 4 | 13.4 GB | Medium | C42 |
Q5_K_MBest for your GPU | 5 | 15.8 GB | High | C44 |
Q6_K | 6 | 18.0 GB | High | C44 |
Q8_0 | 8 | 23.5 GB | Very High | F0 |
F16 | 16 | 45.1 GB | Maximum | F0 |
~$2,499 MSRP