Slotstream Runs 125B Qwen3.8 Flash Next on a 48GB Mac at 12 tok/s Via SSD-Streamed Experts

Developer carloslfu released Slotstream, an OSS runtime that runs the 125B-parameter Qwen3.8-Flash-Next MoE on a 48GB M5 Pro Mac by streaming expert weights from SSD via pread into a shared cache across all 48 layers. Reported figures: 104GB on-disk 4-bit footprint, ~32GB peak RAM, ~12 tok/s warm decode and ~3s cold-start-to-first-token, with Ollama and OpenAI chat API compatibility plus optional speculative decoding.