[GUIDE] Qwen3.8 Flash-Next on a single RTX 5080 + 64GB RAM: >250 tok/s tg and 5700 tok/s prefill in real agent coding sessions (Basalt + new Unsloth UD-IQ3_XXS + Lamina drafter),+ how it compares to llama.cpp

TL;DR Basalt is the engine and it's doing the heavy lifting. Same model file, same box: about 3.5x llama.cpp's decode without drafting and ~5x with it, and ~7x faster prompt reads. Unsloth's new UD-IQ3_XXS (Dynamic v3.1