Thinking Machines open-sources 26B variable-loop model; Inkling 2 Mini beats 594B predecessor at 2.7% size

ChengleiSi · x · 2026-11-21

Thinking Machines Lab released Inkling 2 Mini (aka Loopling), a 26B-parameter looped transformer that runs on a laptop and scores 20.1 on AA Index, outperforming the 594B Inkling 1.0 from July in almost every way at just 2.7% of the size. It also open-sourced (Apache 2.0) a research variant, TML-d24-varloop-research: a 24-layer decoder-only MoE transformer (8 of 128 experts) whose middle 12 layers share weights and run a variable number of loops (1–5 passes) chosen at inference time — 3.4B active params at 1 pass, +1.0B effective per extra pass. Key empirical finding: long-context benchmarks (GraphWalks BFS, MRCR v2) improve with more passes, while knowledge-intensive ones (AA-Omniscience, SimpleQA Verified) do not.

Related event: Thinking Machines Releases 26B Looped Transformer Inkling 2 Mini, Runs on Laptops(115 posts)→

Original post →

More from Models

Models channel →