AnalysisDevelopersSeptember 21, 2026

Splash engine runs Qwen3.8-27B in native 8-bit on Apple Silicon

A Reddit benchmark reports the Splash engine (by Incoai) running Qwen3.8-27B at 37-55 tok/s in native 8-bit on an M5 Pro with 64 GB unified memory. The post also covers 256k context scaling and a "Reasoning Cliff" observed during testing.

1 source

More stories today

Open the live feed