On November 20, 2026, Thinking Machines Lab released Inkling 2 Mini, an open-weights model with 26.4B total parameters and 3.4B active parameters, with a memory footprint of about 20GB — small enough to run on a consumer laptop. Instead of adding more parameters, it runs its middle 12 layers three times for every token, trading compute for reasoning quality.
Three loops, no extra parameters
Inkling 2 Mini has 24 layers: the two middle blocks of six layers each run three times, while the first and last blocks run once, bringing weight activations to 5.4B per token. The lab's reasoning is that the middle layers handle abstract reasoning, so running them more often lets the model think longer without writing all of its reasoning down. A variable-loop research model released the same day backs this up: its no-chain-of-thought time horizon rose from 16 seconds at one pass to 37 seconds at three, and serial reasoning kept improving out to ten loops — beyond the number it was trained with — while the chain of thought grew shorter but stayed legible.
Leading its size class, with a knowledge ceiling
Among open-weights models under 100B parameters, Inkling 2 Mini leads on agentic coding: 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2. It also patches the small-model knowledge gap with tools: 79.0% on BrowseComp and 85.3% on DeepSearchQA, the best scores in its size class. But multi-turn chat and heavily constrained instructions are not its strengths — 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, both behind Qwen3.8-27B — and knowledge-heavy benchmarks barely improve with more loops. In an independent Stanford test of intelligence per joule, it scored 0.703 on an NVIDIA DGX Spark, clearly ahead of the Qwen3.6-35B-A3B and Gemma-4-26B-A4B models tested alongside it.
Most of the weights are trained and deployed natively in NVFP4, with the last four layers in MXFP8, and the model went live on Thinking Machines Lab's own Tinker platform on release day, for inference and customization. The team even ran it offline on a MacBook Pro during a flight and had it build a playable flight simulator in a few minutes. For local deployment and agent development, the real news is that looping works: stronger reasoning without more parameters. What to watch next is how it performs in the community's tests on ordinary laptops, and whether the looping design gets scaled up to larger models. The larger model in the same family was released earlier — see Inkling 2 Small: 594B parameters, only 17B active.