parallelquant
November 21, 2026 · The Decoder

Thinking Machines releases open Inkling 2 Mini, a looped model for laptops

Thinking Machines released Inkling 2 Mini, an open 26.4-billion-parameter mixture-of-experts (MoE) model with 3.4 billion active parameters that runs 12 of its 24 layers three times per token. It fits on a 32GB Mac or a single DGX Spark and scores 79.2% on SWE-bench Verified. On Artificial Analysis's Intelligence Index it ties Qwen3.8 27B at roughly six times lower cost.

Why it matters: Looping layers lets a model 'think' internally without spending output tokens on visible reasoning, attacking the main cost driver of today's reasoning models. If the approach holds up, it shifts competition from parameter count toward compute reuse per token, and it brings near-frontier coding scores onto consumer hardware — the same local-first trend seen in recent RL-trained open coders like DaisyRose-27B.

Related updates