A local AI model gets interesting when the Wi-Fi goes out
Thinking Machines Lab put 20 simulated residents on one offline MacBook. Inkling 2 Mini makes a smaller test worth trying: can an agent finish one real job on your own machine?
In Thinking Machines Lab’s demo, a tiny town keeps going after the MacBook running it goes offline. Its residents plan their days, trade gossip and hold bake-offs. All 20 share the same model. It’s a charming demo, but the part worth stealing is turning off the network.
Thinking Machines released Inkling 2 Mini on November 20. It takes text and images, returns text, and ships as downloadable weights under Apache 2.0. The lab’s offline town ran on a MacBook Pro with an M5 Pro and 48 GB of memory. The original checkpoint is about 20 GB before you account for the memory used while running it. Hugging Face has links to smaller quantized builds and local runners. That makes it easier to try on less capable hardware, not a guarantee that it will be fast enough for a whole coding session.
There’s also a difference between getting a model onto a laptop and getting an agent to finish a job there. A model can run locally while the tools around it still call the web, send logs elsewhere or fall back to a hosted model. And the lab’s own release notes acknowledge weaknesses on multi-turn chat and tasks with many constraints. Twenty simulated villagers are not twenty independent jobs that have to get the details right.
A better first test than an empty chat: pick one task your team repeats, such as explaining a failing test and proposing a small patch in a disposable checkout. Download the model and dependencies first, turn off network access for the test environment, and give the agent a clear stopping point. Check whether it finishes, whether the patch passes the same test and how long the whole thing takes. If it stops because a tool needs the internet, that’s useful too. You’ve learned which part of the workflow still depends on a service.
The promise of a local model isn’t that it should do everything offline. It’s that you can decide which parts need to leave your machine. The way to find that line is to take one job all the way to done.
Sources
- 01 Introducing Inkling 2 MiniThinking Machines Lab
November 20 release, the lab’s offline town demo and its stated multi-turn and constraint-following limitations.
- 02 Inkling 2 Mini model card and checkpointThinking Machines Lab · Hugging Face
Input/output modalities, Apache 2.0 license, approximately 20 GB checkpoint and local hardware guidance.
- 03 Welcome Inkling 2 Mini, Cutting-edge Performance for your Local SetupHugging Face
November 20 launch-partner guide to local runners and quantized builds; not independent performance evidence.
Related Kryden News
When should an AI agent stop and ask first?
Google DeepMind, MIT Technology Review, and Gartner are all pointing at the same practical question: what should an AI helper be allowed to do, and when should it stop for a human yes?
Robot demos are only useful when failure is part of the loop
NVIDIA GEAR's ENPIRE research shows coding agents moving from code generation into robot-training loops, where memory, resets, verification, parallel evaluation, and failure repair matter as much as the base model.