Running Qwen 3.8 Flash Next on nextsycl: 63 GB of Model on a 32 GB Card
nextsycl is a Rust + SYCL runtime for Intel Arc that runs a 63 GB MoE on one 32 GB Arc Pro B70 — the experts that don't fit live in pinned host RAM, the PLE table is streamed from disk, and MTP drafts double the single stream. What it took to get 1,200 tok/s prompt read and ~78 tok/s decode, including the prompt-cache bug that made every agent turn pay a full re-read.