AI on your device
In ElastOS, your AI works at home. A local model answers on your own computer. Nothing you ask leaves it.
Local models
You get models from Marketplace > Models and use them in Assistant.
| Model | Download size | Good for |
|---|---|---|
| SmolLM2 135M Instruct | about 145 MB | Very small and fast. Short, simple answers. |
| Qwen2.5 1.5B Instruct | about 1.1 GB | Better answers. Needs about 3 GB of free memory. |
- Both models are open models under the Apache 2.0 licence.
- They run with llama.cpp, a widely used open-source engine.
- Each model is a signed capsule in a signed catalog. Before your Home keeps a model, the Runtime checks it against its signed catalog entry.
- A short network break during a download is retried by itself. If the download still stops, select Retry.
Hosted models (optional)
On a Mac, an admin can add a Venice or OpenRouter key in System > Models. Hosted models are bigger and faster, but your question goes to that company.
Nothing is sent until you approve it in Inbox.
What models can and cannot do today
Models answer questions in Assistant. They cannot read your files, open websites or send messages. The very small SmolLM2 model sometimes says it did one of these things. It did not.
- On a slow device such as a Jetson, getting Qwen can take about 43 minutes, and its first reply about 22 seconds.
- The model provider is the one provider that is partly confined today.
Why local AI matters
When AI runs on someone else’s servers, they see everything you ask. When it runs on your computer, nobody else does. ElastOS treats AI agents like any other software: they get only the access you grant, through the same checks as a person.
Today in v2: local models on your computer, and optional hosted models that wait for your approval.
Where it is headed: AI agents as sealed capsules with only the access you give them, and sharing models with your contacts. See the roadmap.