Qwen3.8 125B (6B active) Flash-Next

User reports Qwen3.8 Flash-Next at 86.8 output t/s and 1,207 input t/s prefill on two Radeon AI PRO R9700 32GB cards, at 105,884 input tokens. Setup is Strata engine 0.1.56 on Windows with IQ3_S weights and a 196,608-token context setting, on a Ryzen 7 7800X3D with 128 GB DDR5. A real Hermes research run averaged 95 output t/s across 14 responses with input context up to 85,334 tokens. Synthetic long-context tests gave 123 t/s at 98,304 input tokens and 148.7 t/s at 192,512, which the user says do not establish sustained real workload speed.