AnalysisAI ModelsAugust 20, 2026

Qwen 3.8 27B KV cache: F16 vs q8_0 differ in output

A user reports that F16 and q8_0 KV cache quantizations for Qwen 3.8 27B are not equivalent, with F16 producing more careful and detailed outputs in structured and free-form generation. Differences were observed on an AMD R9700 with ROCm.

1 source

More stories today

Open the live feed
Qwen 3.8 27B KV cache: F16 vs q8_0 differ in output