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

A developer spent three weeks testing local Qwen3.8 models using new low-latency SGLang/vLLM recipes, specifically DFlash2 2.8x. The tested builds included RadixArk + Inferact 27B NVFP4, 27B BF16, orcarouter 27B Uncensored, and Flash-Next NVFP4. In a gauntlet test, RadixArk/Qwen3.8-27B-NVFP4 scored 700, outperforming Claude Fable 5.1 and other Qwen3.8 variants. The 27B and Uncensored models beat 12 out of 15 enemies, while Flash-Next beat 13 out of 17.