2× Radeon R9700 for local LLMs
4 measured runs. 32 GB of VRAM; memory measured at 0 GB/s, which sets the roofline for every model below.
| Model | Quant | Engine | tok/s | Roofline | Share |
|---|---|---|---|---|---|
| Qwen3.8-27B | Quark W4A16 | vLLM + MTP k=2 | 92.3 | – | – |
| Qwen3.8-27B | Quark W4A16 | vLLM | 56.1 | – | – |
| Qwen3.8-Flash-Next | GSQ-RCO IQ3_S | llama.cpp + MTP k=3 | 72.5 | – | – |
| Qwen3.8-Flash-Next | GSQ-RCO IQ3_S | llama.cpp | 39.4 | – | – |
Share = measured tok/s over the roofline: the bytes a token reads over this GPU's measured memory speed. Open a quant for the test machine and the per-role and per-kernel breakdown.