Thinking Machines' loop dial tested: 3 inner loops cut written reasoning 41% while MATH accuracy rose to 99%

SonglinYang4 · x · 2026-11-22

Thinking Machines released a research version of Inkling 2 Mini (TML-d24-varloop-research) that lets you choose how many times its middle layers loop.

MohithAchu spent the morning turning that dial across math, code and trivia questions, collecting 5,334 responses. Going from 1 loop to 3 loops cut written reasoning tokens by 41% (1,097 → 642 mean reasoning tokens on 200 MATH problems) while MATH accuracy improved from 92.5% to 99%.

The takeaway: internal recurrent compute can substitute for a chunk of explicit chain-of-thought, trading fewer output tokens for higher accuracy.

Original post →

More from Models

Models channel →