AMD Instinct MI350X 288GBBest value
C71.7 tok/s decode
~$8,000 MSRP
Can it run?
Runs with offload (needs ~16.6 GB host RAM)
Using Q4_K_M in vLLM
Fit status
Runs with offload (needs ~16.6 GB host RAM)
Decode
50.6 tok/s
TTFT
3823 ms
Safe context
15K
Memory
274.7 GB / 256.0 GB
Offload
10%
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Very compromised (needs ~18.8 GB host RAM) | 50.2 tok/s | 5606 ms | 30K |
| Chat | C | Runs with offload (needs ~15.5 GB host RAM) | 50.8 tok/s | 2077 ms | 7K |
| Coding | C | Runs with offload (needs ~16.6 GB host RAM) | 50.6 tok/s | 3823 ms | 15K |
| RAG | C | Very compromised (needs ~18.8 GB host RAM) | 50.2 tok/s | 7007 ms | 30K |
| Reasoning | C | Runs with offload (needs ~16.6 GB host RAM) | 50.6 tok/s |
How Llama 4 Maverick 17B 128E (400B params) fits at each quantization level on AMD Instinct MI325X 256GB (256.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 156.0 GB | Low | C43 |
Q3_K_SBest for your GPU | 3 | 196.0 GB | Low | C44 |
ollama run llama-4-maverick-17b-128ehuggingface-cli download llama-4-maverick-17b-128eUpgrade options
~$8,000 MSRP
| 4518 ms |
| 15K |
| 4 |
224.0 GB |
| Medium |
| C44 |
Q4_K_M | 4 | 244.0 GB | Medium | C44 |
Q5_K_M | 5 | 288.0 GB | High | F0 |
Q6_K | 6 | 328.0 GB | High | F0 |
Q8_0 | 8 | 428.0 GB | Very High | F0 |
F16 | 16 | 820.0 GB | Maximum | F0 |