> ## Documentation Index
> Fetch the complete documentation index at: https://docs.stenox.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Manage local models

> Download, select, load, and remove local Stenox models.

## Download and select

Open **Settings → Models** to browse local model files and download them. Downloading a model in the background does not switch the active selection. Select speech in your [profile](/profiles/creating-profiles), and choose a language model in **Text cleanup → Local** for optional Polish.

Selecting an uncached local speech model starts its download and preparation. A cached model loads from disk. Initial downloads need a network connection; inference then runs on your Mac.

| State | What it means |
| - | - |
| Not downloaded | Model files are not ready on disk. |
| Downloading | Stenox is fetching the files; keep network access and storage available. |
| Downloaded / cached | Files exist locally, but the model may not be in memory. |
| Loading | Stenox is preparing the model for inference. |
| Ready | The selected runtime is loaded and can process input. |

Download labels are estimates from Stenox's catalog. Actual storage includes supporting files, and runtime memory can exceed download size. Running local speech, a meeting diarizer, and meeting insights together uses more memory than speech alone. No fixed per-model RAM minimum is guaranteed here.

## Loading and memory

Stenox prepares the selected local speech model on startup and keeps eligible selected models ready across ordinary idle periods. It restores eligible cached models when needed; changing a selection or disabling local cleanup releases the old runtime through the model lifecycle. Do not expect a universal five-minute unload timer.

Notetaker borrows the active speech model. While capture holds that model, competing jobs and model changes are guarded. Pausing retains meeting state. Ending releases the reservation and meeting-only models while keeping dictation's speech selection ready.

**Polish** waits for a selected model that is already warming up, but does not silently download a missing cleanup model or use another provider. A cleanup failure leaves the dictated text available.

## Remove downloaded files

Use the delete control beside a downloaded model in **Settings → Models**. Read the confirmation, including any profiles that use it. Removal is guarded while a model is loading or reserved. Removing files means they must be downloaded again before use; unloading a runtime only releases memory.

## Catalogs

* [Parakeet](/providers/transcription/parakeet-local) and [Whisper](/providers/transcription/whisperkit-local)
* [Cohere Transcribe](/providers/transcription/cohere-local)
* [Local AI cleanup and legacy models](/providers/llm/mlx-local)
* [Meeting-only processing models](/meetings/processing)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.