Analysis Finds Half of Qwen3.8 Experts Nearly Idle

An analysis of Qwen3.8-Flash-Next using roughly 700k tokens of Claude Code session traces found that half of its 24,576 experts are nearly idle, with just 64 experts carrying about half the compute, suggesting 50% of experts could be pruned.

2026-08-27 ~ 2026-08-27 · 2 related posts

• Qwen 38B MoE Audit: Half of Experts Are 'Passengers' with Minimal Load — EyalToledano · 2026-08-27
• Qwen3.8 Expert Analysis: 50% Can Be Pruned With Minimal Loss — EyalToledano · 2026-08-27