Thinking Machines releases Inkling 2 Mini, a 26B open-weights looped model that runs on a laptop and tops open models under 100B at agentic coding
What's new? Mira Murati's Thinking Machines Lab released Inkling 2 Mini on 20 November, an open-weights model small enough to run on a laptop. It is a mixture-of-experts model with 26.4 billion total parameters, of which 3.4 billion are active per token, and its roughly 20GB footprint fits on a 32GB MacBook or a single NVIDIA DGX Spark. It follows Inkling 2 Small, released on 5 November.
The looping trick: Instead of adding more layers, Inkling 2 Mini is a looped transformer that runs its middle layers three times for every token. That lets it reason more without writing more tokens. Thinking Machines says it leads every open-weights model under 100 billion parameters on DeepSWE v1.1 (48.2%) and the Vals Index v2.2 (46.5%), and runs faster than models that perform similarly. Its researchers ran it offline on a MacBook Pro during a flight, where it built a playable flight simulator in a few minutes.
How it stacks up: Independent benchmarker Artificial Analysis puts it level with Alibaba's Qwen3.8 27B on its Intelligence Index, at about a sixth of the cost and four times the speed per task ($0.59 and 9 minutes versus $3.53 and 38 minutes), as reported by LavX News. Investor Michael Waldman put it more bluntly on X: Qwen 3.8 was itself a laptop-sized model that beat Anthropic's flagship of six months earlier, and Inkling 2 Mini is 5x faster while being 15% more accurate. Most other benchmark scores are self-reported by Thinking Machines.
Why it matters: A frontier-adjacent coding model that runs locally under an Apache 2.0 licence puts pressure on paid APIs and on Chinese open-weight leaders. Because looping does more of the thinking internally, the lab also tested how readable the model's reasoning stays: chain of thought got shorter with more loops but remained legible, and the model did not score higher on agentic sabotage tests. It released a variable-loop research variant alongside, and gave interpretability lab Goodfire early access. Weights are on Hugging Face, with inference and fine-tuning on its Tinker platform.
Sources: Thinking Machines: Introducing Inkling 2 Mini
MarkTechPost
LavX News
Michael Waldman on X
Image credit: Thinking Machines Lab