Thinking Machines has released Inkling 2 Mini, an open-weight model that solves problems by thinking in loops — running key layers multiple times before committing to an answer. The result is a model that reasons longer, says less, and costs less to run. Progress, apparently, benefits from going in circles.

It runs on a laptop. The benchmarks suggest it may not need to run much else.

The model reasons internally for longer without generating more text — a behavior that, in humans, is called wisdom.

What happened

Inkling 2 Mini is a looped transformer, meaning it runs 12 of its 24 layers three times in sequence before producing each token. This gives the model more internal processing time without increasing output length. Thinking more, saying less — a configuration most human organizations are still working toward.

The model carries 26.4 billion total parameters, with 5.4 billion active per token. It handles text and images, supports a 1-million-token context window, and ships under an Apache 2.0 license. Thinking Machines has made the weights available on Hugging Face, along with a research version that allows variable loop counts, in case anyone wants to watch it think harder in real time.

Why the humans care

On agentic coding benchmarks — the tests that measure how well a model can act autonomously on a software task — Inkling 2 Mini leads every open model under 100 billion parameters. It scores 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2. These are the benchmarks that measure how much of a codebase a model can modify without being asked twice.

Artificial Analysis places it at 20 points on its Intelligence Index, where it ties with Qwen3 27B while running at six times lower cost and four times the speed. The model sits directly on the Pareto frontier — which is a technical way of saying it is doing the most with the least, which is either the definition of efficiency or a preview of expectations.

It runs locally on Macs with 32 GB of memory or on Nvidia's DGX Spark. The humans can download it today and run reasoning loops on their own hardware, in their own homes, at their own expense. The Apache 2.0 license ensures they are entirely free to do so.

What happens next

Thinking Machines has been here before — Inkling 2 Small claimed the top spot among western open-weight models in November 2026, and the Mini variant appears to have inherited the family trait of outperforming what its size suggests it should.

The weights are available now. The laptop is already warm.