Back

Thinking Machines Drops Inkling 2 Mini, an AI Model Built to Run on a Laptop

Thinking Machines Drops Inkling 2 Mini, an AI Model Built to Run on a Laptop

Thinking Machines Lab is bringing its AI ambitions down to laptop size ... and this little machine has a few tricks up its sleeve.

The company released Inkling 2 Mini on November 20, pitching the open-weight model as a capable assistant for tasks like agentic coding that can run locally instead of relying on a remote server.

The headline number: roughly 20 GB for the model. But before you go firing it up on just any old laptop, there’s a hardware catch.

Big Brain, Laptop Footprint

According to the company’s model card, the checkpoint can run on Apple silicon devices with at least 32 GB of memory. Other GPUs and platforms are possible, with 32 GB or more of video memory recommended.

Thinking Machines says its researchers put the offline promise to the test on a flight, running the model on a MacBook Pro and having it build a playable flight simulator in a few minutes. Not a bad way to kill time in airplane mode.

The weights are available to download under the Apache 2.0 license. Developers can also access the model for inference and customization through the company’s Tinker platform.

Three Passes, One Token

Here’s where things get nerdy: Inkling 2 Mini is a looped transformer. Its middle layers run three times for each token, allowing the model to reuse parts of its architecture rather than storing a separate set of layers for every pass.

Research firm Goodfire’s architecture analysis explains that the model has 24 layers, with the middle 12 looped three times. Thinking Machines says that amounts to 48 layer applications per token.

It’s also a mixture-of-experts model, with 26.4 billion total parameters and 3.4 billion active. Accounting for the repeated passes, the company reports 5.4 billion weight activations per token.

The Scores ... and the Fine Print

Thinking Machines reports a 48.2% score on the DeepSWE v1.1 coding benchmark and 46.5% on Vals Index v2.2. The company says those results lead reported scores for open-weight models with fewer than 100 billion total parameters. Its announcement also cautions that evaluation setups vary by model.

There are limits, too. The company says Inkling 2 Mini trails Qwen3.8-27B on its tests of multi-turn chat and complicated instructions, and falls behind larger models on some reasoning-heavy benchmarks. Its model card also warns that it can produce plausible but incorrect information.

And the looping raises a safety question: If a model does more work internally and writes less of its reasoning down, how much can an observer see? In experiments with a separately released variable-loop research model, Thinking Machines says written reasoning became shorter as loops increased but remained legible in its tests. That’s a research finding, not a guarantee that every answer will be easy to monitor.

The takeaway? This Mini is aiming to do serious work without a giant hardware footprint ... but it still needs a well-equipped machine and a human willing to check its homework.

Comments

No comments yet. Be the first to comment!

Leave a Comment
Maximum 30 characters
Maximum 100 words

Comments will be visible after approval.