Qwen3.8-Flash-Next on One RTX 5090: Why I Kept IQ4_XS v3.1 at 256K
Rebuilding my workstation from Fedora 44 to Omarchy turned into a quantization experiment. NVFP4 was my first pick; then a fresh Unsloth release changed the candidate. With both remeasured on Strata 0.1.57, IQ4_XS v3.1 tied the core checks, passed one focused coding rerun NVFP4 failed, decoded about 21% faster on short prompts, and left nearly 8 GiB more system RAM available. The account includes the exact settings, the limits of the tests, and why I kept native 256K instead of declaring victory at 512K.