Thinking Machines Lab researchers put a laptop in airplane mode mid-flight and asked their new model for a flight simulator. In the lab’s Inkling 2 Mini launch post, the model builds a playable one in a few minutes, offline, on a MacBook Pro with 48GB of memory.

That is the pitch for Inkling 2 Mini, released on 20 November: a coding and agent model for your own machine. In a second demo, one offline MacBook runs all 20 villagers of a pixel town the model also built.

What you need to run it

The Inkling 2 Mini model card on Hugging Face puts the checkpoint at about 20GB. It supports Apple silicon Macs with 32GB or more of memory and NVIDIA’s DGX Spark desktop. On other GPUs, 32GB of video memory is recommended.

The weights are open under the Apache 2.0 license, which allows commercial use. On a Mac, the Ollama library page for Inkling 2 Mini gets it running with one command: ollama run inkling-2-mini. It reads text and images and has a context window (how much it can read at once) of up to 1 million tokens.

For scale, Tencent’s Hy4, which we covered on Friday, needs about 944 GB of GPU memory for its open weights.

The trick: it thinks in loops

Inkling 2 Mini has 24 layers, and the middle 12 run three times for every token, reusing the same weights. So it thinks more per word without storing more parameters.

It is also a mixture of experts model, so each token only wakes a few specialist sub-networks. Of 26.4 billion parameters, 3.4 billion are active per token, or 5.4 billion with the loops counted. Our explainer on which language models run on a laptop covers why active parameters set speed and total parameters set memory.

The lab flags a catch: a model that reasons inside its layers writes less reasoning down for humans to check. In its own tests, that written reasoning got shorter with more loops but stayed readable.

What the lab claims, and what outsiders found

Thinking Machines says Inkling 2 Mini leads every open model under 100 billion parameters on coding tests such as DeepSWE, where it reports 48.2%. Those are the lab’s own scores.

Artificial Analysis runs the same tests on every model. Its independent review of Inkling 2 Mini scores it 20 on its Intelligence Index, level with Alibaba’s Qwen3.8 27B. Each task cost $0.59 and took 9 minutes, against $3.53 and 38.2 minutes for Qwen. It did use about 38% more output tokens for the same score.

Against bigger models, it is small. The Vals AI results for Inkling 2 Mini show 46.46% and 38th place of 43 models on the Vals Index. Our guide to the big AI leaderboards explains why such rankings differ.

Hugging Face engineers tried it informally in their post on running Inkling 2 Mini locally. A 4-bit version built a website for a joke startup that matches house plants with therapists, graphics included. They call it at least comparable with Qwen 3.8 27B and say more validation is needed.

Where it slips

The model card is blunt: Inkling 2 Mini is more likely than larger models to state wrong facts with confidence, and it can get stuck in long reasoning loops while debugging. With tools, it may delete or overwrite files without asking and misreport what it did. The lab adds that it is not the strongest small model for long, multi-turn chat.

What to do this week

If you have a Mac with 32GB or more, install Ollama and give Inkling 2 Mini one small coding job, such as a single-page app. Use an empty folder, cap its output and agent turns as the model card advises, and compare the result with your usual cloud model.

Sources

Lars Strother

Lars Strother is the senior editor for models and hardware at AI Observer, covering models and research, chips and hardware, and robotics. He came to journalism from engineering and still reads the benchmark tables before the press release. His coverage is measured and skeptical, and he wants to know what was tested, on what, and against which baseline.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.