Thinking Machines' 26B looped-MoE Inklings 2 Mini beats GLM-5.2 and Opus 4.5 on benchmarks

ziqiao_ma · x · 2026-11-21

Thinking Machines Lab released Inklings 2 Mini: a looped-MoE model with 26.4B total params, only 4B active per token, running in 20 GB — yet it beats GLM-5.2 Full (753B, 40B active, 28x larger) and Claude Opus 4.5 on many benchmarks: Vibe Code Bench 77.9 vs 64.0, MCP Atlas 82.7, BrowseComp 79.0, DeepSWE 74.9, SWE-bench Verified 79.2 (GPT-5.2 Thinking: 80.9). The looped MoE approach (reusing a small active set across iterations) shows remarkable efficiency, though it still trails GPT-5.2 on ARC-AGI v2 and Automation Bench.

Original post →

More from Models

Models channel →