Swift 1.5 matches Qwen3.8 27B quality on M5 Max while writing 34% fewer tokens

DerTomsn · reddit · 2026-09-30

A Redditor benchmarked Swift-1.5-Qwen3.8-27b (oQ8e, MLX) on an Apple M5 Max 64GB via llm-bench.io with 262k context and thinking at xhigh, against base Qwen3.8 27B.

Bottom line: equal quality, roughly a third faster thanks to leaner reasoning. The author plans to use it as a daily driver. All eight raw benchmark runs are linked.

Original post →

More from Infra

Infra channel →