Skip to content
Predictive Systems
PSI Daily

A Strong Open Model You Can Run Fully Offline

SovereigntyAllan C. Tan, MS

Thinking Machines Lab released Inkling 2 Mini, an Apache 2.0 open model of about 20GB that can run fully offline on a well-equipped laptop. The lab's coding scores have not been independently checked.

AI illustration: a walking laptop carrying a large brain with amber light looping around it, with overlay: A Strong Open Model You Can Run Fully Offline

Thinking Machines Lab, the company founded by Mira Murati, released Inkling 2 Mini on 20 November. It is an open-weight model under the Apache 2.0 license, which allows reuse, including for commercial work. The download is about 20GB, small enough to run on a well-equipped laptop, and it can work fully offline. That is a form of sovereignty you can download. You keep the weights, you run the model on your own machine, and you do not need someone else's cloud to use it.

A compact design, with a loop in the middle

Inkling 2 Mini is a mixture-of-experts model. It stores many specialist parts, but for each word it writes it turns on only a few of them. The full set has 26.4 billion parameters. Only 3.4 billion of those run at once. The middle layers also run three times for every word, reusing the same weights. The lab says this looping lets the model think more carefully without writing a longer explanation. With the loops counted, about 5.4 billion parameters are active for each word.

The model card says it needs a reasonably powerful machine: an Apple silicon computer with 32GB of memory or more, or another graphics processor with at least 32GB of video memory. The lab showed the model running offline on a MacBook Pro during a flight, where it built a playable flight simulator. That demonstration is the company's own account.

The coding scores are still the lab's

The lab says Inkling 2 Mini beats every open model under 100 billion parameters on coding tests, including a score of 48.2% on DeepSWE v1.1. That is a company claim, and the lab notes that test setups vary by model. Hugging Face staff, in informal tests, said the model was at least comparable with Qwen3.8-27B, and that more validation is needed. Artificial Analysis, which runs its own tests, gave the model 20 points on its Intelligence Index and said its answers were very wordy. Early testers on Hacker News, where the release was a leading story, also found that simple coding tasks often worked, but that the model sometimes got stuck in long reasoning loops. The lab's own model card warns about the same habit.

The lab also says looping keeps the model's written reasoning short but still readable, which would help people who want to check what the model is thinking. Those tests used a research version that can change how many loops it runs. The claim has not been independently checked. Looping also moves more of the work inside the model, so less of the thinking may appear in the text at all.

Whether the coding scores hold up will take more independent tests. What is already clear is the offer: a capable open model, under a permissive license, that you can keep and run without sending data to someone else's server.

Sources

  1. Introducing Inkling 2 Mini (Thinking Machines Lab, 20 November 2026)
  2. Inkling 2 Mini model card (Thinking Machines Lab, Hugging Face, 20 November 2026)
  3. Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup (Pedro Cuenca, Aritra Roy Gosthipaty and Guarin Flück, Hugging Face, 20 November 2026)
  4. Inkling 2 Mini (Hacker News thread, 20 November 2026)
  5. Inkling 2 Mini: Intelligence, Performance and Price Analysis (Artificial Analysis)
  6. New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop (Matthias Bastian, The Decoder, 21 November 2026)

Illustration: AI-generated for PSI.