M5 MacBook Air with 24GB runs Qwen3.8-Flash-Next at 49 t/s locally

On October 25, 2026, a Reddit user posting in the LocalLLaMA subreddit reported running the quantized Qwen3.8-Flash-Next (ISTA GSQ-RCO Q2_0) on a MacBook Air M5 with 24 GB of memory, using a lightly modified fork of Splash 1.4.0 together with the Flash-Next-DFlash2 draft model. Because the full model did not fit in RAM, the setup streams expert weights from the SSD on demand: of the 24,576 expert slots in total, 6,800 (~28%) are held on disk in roughly 8.8 GB, and the rest is read as needed. The configuration retains a 256K context. In the user's own measurements, adding cache-aware routing (bias 2, with the top 4 tokens protected) raised average decode throughput across code, tool-calling, and explanation tasks from 28.8 t/s to 43.1 t/s. Breaking that down by task, the table lists code at 31.6 t/s rising to 48.6 t/s, tool calling at 29.5 t/s rising to 45.7 t/s, and explanation at 25.4 t/s rising to 35.1 t/s. As of now, this single self-reported Reddit post is the only account of the experiment, and there is no independent confirmation or official statement from the model or runtime developers.