I spent 3 weeks testing local Qwen3.8 on the new low-latency SGLang/vLLM recipes: DFlash2 2.8x. Builds: RadixArk + Inferact 27B NVFP4, 27B BF16, orcarouter 27B Uncensored, Flash-Next NVFP4

Hey guys, Last time I tested Qwen3.8-Flash-Next on its own. This time I put three Qwen3.8 checkpoints through the same 10 tests on the same RTX PRO 6000: RadixArk/Qwen3.8-27B-NVFP4 (dense) orcarouter/Qwen3.8-27B-Uncensored-NVFP4 (dense, uncensored fine-tune) RadixArk/Qwen3.8-Flash-Next-NVFP4 (MoE) E