Qwen3.8-27B Benchmarked on AMD R9700: Up to 227 tok/s

samsja19 · x · 2026-08-26

Benchmarks show that Qwen3.8-27B, running the Unsloth UD-IQ4XS quantization on an AMD R9700 GPU, achieves up to 227 tokens per second. Compared to a Q80 reference, the quantized version maintains 8-bit-class performance with a mean KL divergence of 0.018 and 94% top-1 agreement.

Original post →

More from Models

Models channel →