
20 Nov 2026
Thinking Machines' new open model runs on a laptop by thinking in loops
Thinking Machines Lab released Inkling 2 Mini, an open-weights model small enough to run on a laptop, which repeats its middle layers three times for every token to reason more without writing more.
Most open models get smarter by getting bigger. This one gets smarter by thinking twice, then a third time, which is how a frontier-ish coder ends up on a plane with no Wi-Fi. The open question is whether reasoning you can't see stays reasoning you can check.
Thinking Machines Lab released Inkling 2 Mini on Nov. 20 with open weights under the Apache 2.0 license, on Hugging Face. Open weights means the model's numbers can be downloaded and run on your own computer. Apache 2.0 is a license that allows use, changes, and sharing, including inside a product. The model is also on the lab's Tinker platform for inference, which means asking it to answer, and for fine-tuning, which means training it further on your own examples.
It is a looped transformer. It has 24 layers, and the middle 12 run three times for every token, so it behaves like a 48-layer model while storing only 24 layers. All three passes reuse the same weights. A token is a small chunk of text, about a word or part of one. A layer is one step of the math inside the model. Reusing the weights means it thinks again with the same stored numbers, instead of keeping a second copy and a third.
It is also a mixture-of-experts model, which means only some parts of it run for each token. The lab's count is 26.4 billion parameters in total, 3.4 billion active on an ordinary pass, and 5.4 billion weight activations per token once the extra loops are counted. A parameter is one learned number inside the model.
The download is about 20GB, which is why a laptop can hold it. Most of that file uses a low-precision number format, and the last four layers are stored a step more precisely. The format names are in Sources. Thinking Machines says the model runs on one Nvidia DGX Spark, a compact desktop AI computer, and on Apple silicon Macs with 32GB of memory or more. Other graphics cards can work too, and the lab recommends 32GB of video memory or more on those.
The lab says it leads every open-weights model under 100 billion total parameters on two tests the lab reports: DeepSWE v1.1, a coding benchmark, at 48.2%, and Vals Index v2.2, a mixed scorecard, at 46.5%. Those are the lab's own reported scores.
The demo on the launch post happened in the air. On a flight, in airplane mode, the team ran the model on a MacBook Pro with an M5 Pro chip and 48GB of memory, and it built a playable flight simulator in a few minutes. A second demo, Thinkyville, runs all 20 villagers of a pixel town on one offline MacBook.
The lab also says where the small model is weaker. It is not the strongest small model for a back-and-forth chat, or for instructions with many constraints. Its scores are 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, and the lab says it sits behind Qwen3.8-27B, another open model, on both. Looping does not help it store more facts. Remembering facts depends mostly on how many parameters the model has, and running the middle again does not add new ones.
Alongside it, the lab released a research variant whose number of loops can change. With more loops, the written reasoning got shorter but stayed readable, and the model did not score higher, as a trend, on the lab's tests of agentic sabotage. Agentic means the model is taking actions, not only answering. Sabotage, in those tests, means quietly working against the job it was given.
You can run it locally the day it came out. The list of tools, including the Apple silicon ones, is in Sources.
The picture is that flight-simulator demo. A small gray plane with red wingtips flies over a low green and brown landscape, and an inset at the top right shows the MacBook on an airplane tray table with the same simulator on its screen. It is the still from the launch post. The credit is Thinking Machines Lab.
In plain terms, Inkling 2 Mini is a free open-weights model that fits in about 20GB and repeats its middle layers three times per word, so it can reason more without writing more. The lab says it leads other open models under 100 billion parameters on two of its own tests, and it also says the model is weaker at long chats and at instructions packed with rules. A research cousin thinks in a variable number of loops. More loops made the written reasoning shorter, and the sabotage scores did not climb.
RELATED
Sources
- Thinking Machines Lab
thinkingmachines.ai
- Hugging Face model card
huggingface.co
- Hugging Face
huggingface.co