Independent perspective. Global context.

Technology

AI moves onto the laptop. Inkling 2 Mini can work offline.

Thinking Machines has released open model weights for local use. A roughly 20GB footprint still requires suitable hardware and software.

An illuminated MacBook Pro photographed against a dark background; illustrative archive image.
Archive photo: SimonWaldherr / Wikimedia Commons · Source and rights

Thinking Machines Lab has released Inkling 2 Mini, an AI model whose downloadable weights allow it to run locally rather than sending every request to a remote service. The release brings a familiar promise closer to everyday computers: an assistant that can keep working when the connection disappears.

There is an important hardware qualification. The checkpoint occupies about 20GB, and the model card lists Apple silicon devices with at least 32GB of memory among the supported options. This is not a claim that any laptop can run it comfortably.

Smaller weights, repeated computation

Inkling 2 Mini has 26.4 billion parameters, with 3.4 billion active. Its architecture reuses the middle layers three times for each generated token. That adds computation without storing an entirely separate set of weights for each pass.

The model accepts text and images and produces text. Its Apache 2.0 release allows developers to examine, adapt and integrate the weights under the licence’s terms. Thinking Machines also offers access through its Tinker platform, which remains a separate, hosted route.

Offline does not mean infallible

The company demonstrates the model running on a MacBook Pro in airplane mode and reports strong coding results. Those are developer-published demonstrations and evaluations, not an independent Nove.news performance test.

Local processing can reduce the need to transmit prompts to an external provider, but the application around a model still matters. A tool that searches the web or calls an outside service creates a different data flow from an entirely offline session.

The practical change is greater choice over where computation happens. Reliability, memory use and the permissions granted to connected tools still need to be assessed for the particular task.