GPUs & Macs › AMD Instinct MI25
AMD Instinct MI25: local LLM speed
Llama 2 7B (Q4_0) generates 63.9 tokens/s on the AMD Instinct MI25 (ROCm), and processes prompts at 410 tokens/s.
Measured results
| Model | Generate tok/s | Prompt tok/s | Quant | Backend | Test | Source |
|---|---|---|---|---|---|---|
| Llama 2 7B | 63.9 | 410 | Q4_0 | ROCm FA off | tg128 | llama.cpp #15021 details |
| Llama 2 7B | 54.7 | 439 | Q4_0 | Vulkan FA off | tg128 | llama.cpp #10879 details |
What fits on the AMD Instinct MI25
Estimates for this chip are scaled from a measured run: Llama 2 7B Q4_0 at 63.9 tok/s (ROCm, source). Fit assumes the whole model in GPU memory with ~8K context.
| Model | Best quant that fits | File | Expected tok/s |
|---|---|---|---|
| Qwen3.8-27B | UD-IQ3_XXS | 10.9 GB | 22.4 estimate |
| gemma-4-12B-it | Q8_0 | 12.7 GB | 19.3 estimate |
| Qwen3.5-9B | Q8_0 | 9.5 GB | 25.7 estimate |
| Qwen3.5-4B | Q8_0 | 4.5 GB | 54.5 estimate |
| Qwen3.6-35B-A3B | UD-IQ3_XXS | 13.2 GB | 108 estimate |
| gemma-4-26B-A4B-it | UD-IQ4_XS | 13.6 GB | 58.4 estimate |
| MiMo-V2.6-Distill-Qwen-9B | Q8_0 | 9.5 GB | 25.6 estimate |
| gemma-4-E4B-it | Q8_0 | 8.2 GB | 29.8 estimate |
| Muse-Glimmer-30B | UD-IQ3_XXS | 13.1 GB | 18.6 estimate |
| Qwen3-4B | Q8_0 | 4.3 GB | 57.1 estimate |
| gemma-4-31B-it | UD-IQ3_XXS | 11.8 GB | 20.6 estimate |
| gemma-4-12B-it-qat-q4_0-unquantized | UD-Q4_K_XL | 6.7 GB | 36.4 estimate |
| gemma-3-1b-it | Q8_0 | 1.1 GB | 204 estimate |
| gemma-4-E2B-it | Q8_0 | 5.0 GB | 49.2 estimate |
Similar hardware
AMD Instinct MI50 · AMD Instinct MI60 · AMD Instinct MI100 · AMD Instinct MI210 · AMD Instinct MI300X