Osegi
AI·2 min read

Inkling 2 Mini: a 26B open AI model that runs on a laptop

Thinking Machines Lab had Inkling 2 Mini write a playable flight simulator on a MacBook Pro with the Wi-Fi off. The open model fits in under 20 GB.

A MacBook Pro laptop running an AI-generated flight simulator offline on a plane tray table.
Image: Frandroid

To prove Inkling 2 Mini needs no internet connection, the Thinking Machines team took it on a plane. On a MacBook Pro with the Wi-Fi off, the model wrote a playable flight simulator in a few minutes, the startup says.

Thinking Machines Lab, founded in 2025 by former OpenAI chief technology officer Mira Murati, released the model on 20 November. That is two weeks after the much bigger Inkling 2 Small, which has 594 billion parameters. Some users found that too large, so the company built Inkling 2 Mini with 26 billion. It is aimed at coding and AI agents that chain actions on their own, and the weights are open under the Apache 2.0 licence, commercial use included.

How does Inkling 2 Mini fit in under 20 GB?

Three tricks. First, it was trained natively in NVFP4, a 4-bit format from Nvidia, with only the last four layers kept at 8 bits. A 16-bit version would weigh over 50 GB.

Second, it is a mixture-of-experts model: each layer has 128 specialised sub-networks, and only 8 work on each token (a fragment of a word). So just 3.4 billion parameters are active at any moment. The team started with 384 experts per layer, cut the least-used two thirds, and retrained the rest.

Third, it loops. Of its 24 layers, the 12 middle ones run three times for each token, giving the reasoning depth of a 48-layer model without extra memory. The company says this saves 140% of compute on complex agent tasks.

How good is it?

On DeepSWE v1.1, which tests whether an AI can fix real software bugs alone, Inkling 2 Mini solves 48.2%, against 42.2% for Alibaba's Qwen3.8-27B. These are the lab's own figures and have not been independently checked.

Qwen3.8 wins elsewhere. It is better in everyday conversation and invents less: 71.4% reliable answers on AA-Omniscience, against 63.4%.

What hardware do you need?

The 4-bit version recommended by Unsloth takes 17.6 GB on disk, plus room for context. Suitable machines include a Mac with 32 GB of unified memory, an AMD Ryzen AI Max+ 395 PC, a 32 GB Intel Panther Lake laptop, or a graphics card with 24 GB of VRAM. Unsloth says 16 GB of RAM can work with heavier compression, at a small cost in accuracy. You can run it with Ollama, LM Studio, Unsloth Desktop or llama.cpp (support still experimental), with files from Hugging Face.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:thinking machines labinkling 2 minimira muratilocal aiopen sourceqwen

More from this day

See all →

More in: AI

See all →