Suppose you have a folder full of old project notes. You want an AI assistant to find the decisions you made last spring and turn them into a short summary. You would rather leave the files on your laptop.
That is possible with a suitable local setup. But a download button labeled “local AI” does not tell you everything you need to know.
What changed this week?
On October 13, Liquid AI released LFM2.5-24B-A2B, a model aimed at running AI agents on personal hardware. Its weights—the numerical settings learned during training—are available to download. Liquid also provides a guide to connecting a local model to agent software, which can give it tools for tasks such as reading files.
The release is a reason to look at an increasingly practical question: does this particular job need a remote AI service?
Think of an assistant sitting beside your filing cabinet
With a cloud model, your request travels to another computer, where the answer is generated. With a local model, that computation happens on your own machine. The model and supporting software need to be installed first.
For our folder example, the app would need to read the documents and pass relevant text to the local model. If those steps all happen on the laptop, the text need not leave it to generate a summary. That is a possible design, not a promise about every app that advertises local AI.
A smaller workload does not mean a tiny download
The new model has 24 billion parameters in total, with about 2 billion active for each generated token—a small chunk of text. That distinction explains the “24B-A2B” in its name.
In a mixture-of-experts model, a routing mechanism chooses a subset of internal networks for each token. Imagine a large workshop where only a few workbenches are busy at once. Fewer benches working does not make the whole workshop disappear.
The practical consequence: active parameter count alone cannot tell you whether a model fits your computer. The full weights still need to be stored and made available, and the running software and conversation use memory too. Check the requirements for the specific download and setup.
The model can stay home while a tool goes online
An agent is a model connected to software that can take actions. Ask that agent to search the web or send an email, and part of the task necessarily reaches another service. Local text generation does not keep those actions offline.
Ollama offers a useful example of this distinction: its official FAQ describes both local and cloud models, plus an option to disable its cloud models and web search. Those settings apply to Ollama; a separate app or tool still needs its own check.
For the folder task, ask three concrete questions: Where is the model running? Where are the documents processed? Which connected tools can send information elsewhere? “It runs locally” answers only the first.
A useful first experiment
Start with a few invented notes whose contents you know. Ask for a summary, then check whether the assistant preserved the decisions and omitted anything important. After installation, trying the task with networking disabled can reveal a dependence on online services. Passing that test does not establish what the app might transmit when networking returns.
For someone evaluating local AI, a small task done accurately on their own hardware is more informative than an impressive headline benchmark. The question to carry into the next release is simple: which part of your work can this actually help you do?