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

User compares Qwen3.8 Flash-Next 4-bit against Qwen3.8 27B FP8 on a coding task, both run on 2x R9700 AI PRO 32GB (64GB total VRAM) with vLLM - Radiance. Flash-Next averaged ~2000 prefill and ~150 generation t/s at 262k context and finished in about 30 minutes, but produced a messy, unusable WebUI. The 27B FP8 averaged ~1500 prefill and ~100 generation t/s and took about 60 minutes, but produced a mostly working result. User notes the speeds are real work averages, not benchmarks, and concludes a dense model outperforms the MoE on the same hardware.