# AI models
URL: https://jarvis.ceo/docs/pantry/models
Status: beta

The models behind Jarvis: on-device transcription for the ears, and a brain you choose (Claude, OpenAI, Gemini, or a local model). What runs locally, and what needs your own key.

> [beta] Choosing the brain model is part of Jarvis, in early access. Download the beta, or join the waitlist for full access.

Jarvis uses two kinds of model: the ears that turn your speech into text, and the brain that reasons and acts.

## The ears: transcription

Speech-to-text runs on-device by default, with no internet needed. Smaller models are faster, larger ones more accurate, and you pick the trade-off. If you want a cloud transcription service for extra accuracy, plug in your own key.

## The brain: reasoning and action

For the agent work, choose the model that does the thinking.

- Cloud: Claude, OpenAI, or Gemini, using your own API key.
- Local: run an on-device model for fully offline, fully private operation.

## Picking a trade-off

- Local models keep everything on your Mac and work offline.
- Cloud models are more capable for hard, multi-step work.
- You can mix: on-device transcription with a cloud brain, or all local.

> [privacy] When you use a cloud model with your own key, the request goes from your Mac to that provider, not through us.

Related: Transcription and offline (https://jarvis.ceo/docs/pantry/transcription-and-offline), Privacy and your data (https://jarvis.ceo/docs/pantry/privacy)