> ## 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.

# MLX (Local)

> 100% private, offline AI enhancement with Apple MLX

"og:title": "Stenox Docs"

<Warning>
  This is an **AI ENHANCEMENT** provider. It improves transcribed text with grammar correction and formatting.
  For speech-to-text transcription, see [Transcription Providers](/providers/overview#transcription-providers).
</Warning>

## Overview

MLX is Apple's machine learning framework optimized for Apple Silicon. Stenox uses MLX to run small, efficient language models locally on your Mac for completely private text enhancement.

**Key Benefits:**

* 🔒 **100% Private** - Text never leaves your Mac
* 🌐 **Offline** - Works without internet connection
* 💰 **Free** - No API keys, no usage costs
* 🚀 **Optimized** - Built for Apple Silicon

## Privacy & Security

MLX processes everything locally on your Mac:

* Transcribed text is enhanced on-device
* No data sent to cloud services
* No internet connection required
* No API keys or accounts needed
* Perfect for sensitive or confidential content

<Tip>
  Combine MLX with [WhisperKit (local transcription)](/providers/transcription/whisperkit-local) for **100% private, offline** voice dictation.
</Tip>

## No API Key Required

Unlike cloud providers, MLX requires **no setup**:

1. Select MLX in Stenox Settings
2. Download your preferred model
3. Start enhancing immediately

No registration, no API keys, no authentication.

## Available Models

MLX offers several small language models optimized for Apple Silicon:

| Model             | Size     | Speed   | Quality   | Recommended For                |
| ----------------- | -------- | ------- | --------- | ------------------------------ |
| **Qwen 2.5 1.5B** | \~1 GB   | Fastest | Good      | Quick enhancement, limited RAM |
| **Gemma 3n 2B**   | \~1.2 GB | Fast    | Excellent | Best accuracy, recommended     |
| **Phi-3 Mini**    | \~2.4 GB | Medium  | Very Good | Balanced quality and speed     |
| **Gemma 3n 4B**   | \~2.5 GB | Slower  | Very Good | Higher quality, more RAM       |
| **Qwen 3 4B**     | \~2.5 GB | Slower  | Good      | Alternative to Gemma 3n 4B     |

<Note>
  **Recommended for beginners:** **Gemma 3n 2B** - Best accuracy with fast processing and reasonable size.
</Note>

### Model Selection Guide

<AccordionGroup>
  <Accordion title="Qwen 2.5 1.5B - Fastest and smallest" icon="gauge-high">
    **Best for:**

    * Quick enhancement (2-3 seconds)
    * 8GB RAM Macs
    * Everyday grammar correction
    * Speed over maximum quality

    **Processing time:** \~2-3 seconds per paragraph

    **Quality:** Good grammar correction, basic formatting
  </Accordion>

  <Accordion title="Gemma 3n 2B - Best accuracy (Recommended)" icon="crown">
    **Best for:**

    * Best accuracy for its size
    * 8GB+ RAM Macs
    * Professional writing
    * Best balance of speed, size, and quality

    **Processing time:** \~2-3 seconds per paragraph

    **Quality:** Excellent grammar, great formatting and tone
  </Accordion>

  <Accordion title="Phi-3 Mini - Balanced quality" icon="medal">
    **Best for:**

    * Higher quality enhancement
    * 12GB+ RAM Macs
    * More complex edits
    * Good balance of speed and quality

    **Processing time:** \~3-4 seconds per paragraph

    **Quality:** Very good grammar, reliable formatting
  </Accordion>

  <Accordion title="Gemma 3n 4B / Qwen 3 4B - Larger models" icon="memory">
    **Best for:**

    * Maximum local quality
    * 16GB+ RAM Macs
    * Critical writing
    * Quality over speed

    **Processing time:** \~4-6 seconds per paragraph

    **Quality:** Near-cloud quality, excellent at complex edits
  </Accordion>
</AccordionGroup>

## Setup Instructions

<Steps>
  <Step title="Open Stenox Settings">
    Click the Stenox icon in your menu bar and select **Settings**.
  </Step>

  <Step title="Navigate to Models tab">
    Go to the **Models** tab in the Settings window.
  </Step>

  <Step title="Select MLX">
    Under **AI Enhancement Provider**, select **MLX** from the dropdown.
  </Step>

  <Step title="Choose a model">
    Select your preferred model:

    * **Gemma 3n 2B** (recommended - best accuracy)
    * **Qwen 2.5 1.5B** (fastest, smallest)
    * **Phi-3 Mini** (balanced quality)
    * **Gemma 3n 4B** or **Qwen 3 4B** (maximum quality)
  </Step>

  <Step title="Download the model">
    Click **Download** and wait for the model to download.

    **Download sizes:**

    * Qwen 2.5 1.5B: \~1 GB
    * Gemma 3n 2B: \~1.2 GB
    * Phi-3 Mini: \~2.4 GB
    * Gemma 3n 4B / Qwen 3 4B: \~2.5 GB

    Models are stored in `~/stenox-models/mlx/` by default.
  </Step>

  <Step title="Start enhancing">
    Once download completes, your transcriptions will automatically be enhanced!
  </Step>
</Steps>

## Performance

MLX performance depends on your Mac's hardware and chosen model:

### Apple Silicon Macs

**Excellent performance:**

* Qwen 1.5B: 2-3 seconds per paragraph
* Gemma 2B: 3-4 seconds per paragraph
* Phi-3 3.8B: 5-7 seconds per paragraph

**Optimized with:**

* Metal GPU acceleration
* Apple Neural Engine (ANE)
* Unified memory architecture

### Intel Macs

**Not supported:**

* MLX requires Apple Silicon
* Intel Mac users should use cloud providers ([Gemini](/providers/llm/google-gemini) or [Groq](/providers/llm/groq-llms))

<Warning>
  MLX models only work on Apple Silicon Macs. Intel Mac users must use cloud enhancement providers.
</Warning>

## Apple Silicon Required

MLX models require Apple Silicon (M1/M2/M3/M4). Intel Mac users should use cloud providers like [Gemini](/providers/llm/google-gemini) or [Groq](/providers/llm/groq-llms).

## What MLX Enhancement Does

### Grammar Correction

* Fixes grammatical errors
* Subject-verb agreement
* Tense corrections
* Article usage (a, an, the)

**Example:**

* **Before:** "me and sarah was going to store"
* **After:** "Sarah and I were going to the store."

### Capitalization

* Sentence beginnings
* Proper nouns (names, places)
* Acronyms and abbreviations

**Example:**

* **Before:** "john lives in new york city and works for ibm"
* **After:** "John lives in New York City and works for IBM."

### Punctuation

* Adds missing commas, periods
* Fixes run-on sentences
* Improves readability

**Example:**

* **Before:** "hello how are you doing today its nice to meet you"
* **After:** "Hello, how are you doing today? It's nice to meet you."

### Light Formatting

* Number formatting (spelled out → digits)
* Date and time formatting
* Basic list formatting

<Note>
  MLX models provide **basic to good** enhancement. For maximum quality, consider cloud providers like [Google Gemini](/providers/llm/google-gemini).
</Note>

## When to Use MLX

<CardGroup cols={2}>
  <Card title="Privacy is critical" icon="lock">
    Healthcare, legal, financial, or sensitive content.
  </Card>

  <Card title="Working offline" icon="wifi-slash">
    Airplanes, remote locations, or unreliable internet.
  </Card>

  <Card title="No API costs" icon="dollar-sign">
    No per-use charges. Processing happens on your Mac.
  </Card>

  <Card title="100% local setup" icon="computer">
    Combine with WhisperKit for completely offline dictation.
  </Card>
</CardGroup>

## When to Use Cloud Instead

Consider cloud providers if you need:

* **Faster processing** - Cloud is 2-3x faster (\< 1 second vs 2-5 seconds)
* **Better quality** - Gemini 2.5 Flash and Groq LLMs often outperform small MLX models
* **Intel Mac** - MLX requires Apple Silicon
* **Complex edits** - Tone adjustment, style changes, formatting

<Tip>
  Use **both**! Create different [Profiles](/profiles/creating-profiles):

  * MLX for personal notes (privacy)
  * Gemini for work emails (quality)
  * Auto-switch based on application
</Tip>

## Custom Prompts

Configure how MLX enhances your text in Profile settings:

<Tabs>
  <Tab title="Default">
    ```
    Fix grammar and punctuation errors.
    Keep the original meaning and tone.
    ```
  </Tab>

  <Tab title="Professional">
    ```
    Fix grammar and rewrite in a professional tone.
    Suitable for business emails.
    ```
  </Tab>

  <Tab title="Minimal">
    ```
    Only fix obvious grammar errors.
    Change as little as possible.
    ```
  </Tab>
</Tabs>

## Troubleshooting

<AccordionGroup>
  <Accordion title="Model download fails" icon="cloud-arrow-down">
    * Check internet connection
    * Ensure sufficient disk space (\~2-4 GB per model)
    * Try downloading again
    * Check `~/stenox-models/mlx/` for partial downloads and delete them
  </Accordion>

  <Accordion title="Enhancement is very slow (> 10 seconds)" icon="hourglass">
    * Try a smaller model (Qwen 2.5 1.5B or Gemma 3n 2B instead of larger models)
    * Close other intensive applications
    * Check Activity Monitor for available RAM
    * Your Mac may not have enough RAM for larger models
  </Accordion>

  <Accordion title="MLX option not available / grayed out" icon="ban">
    * You're on an Intel Mac (MLX requires Apple Silicon)
    * Use [Google Gemini](/providers/llm/google-gemini) or [Groq LLMs](/providers/llm/groq-llms) instead
  </Accordion>

  <Accordion title="Enhancement quality is poor" icon="spell-check">
    * Try Gemma 3n 2B (best accuracy) or a larger model
    * Adjust custom prompt in Profile settings
    * For maximum quality, use cloud providers ([Gemini](/providers/llm/google-gemini))
    * Add specific formatting instructions to your custom prompt
  </Accordion>

  <Accordion title="App crashes or runs out of memory" icon="memory">
    * Your model is too large for available RAM
    * Use Qwen 2.5 1.5B or Gemma 3n 2B on 8GB Macs
    * Upgrade to 16GB+ RAM for larger models
    * Close other applications to free memory
  </Accordion>
</AccordionGroup>

## Model Comparison

| Model             | RAM Required | Speed   | Quality   | Best Use Case              |
| ----------------- | ------------ | ------- | --------- | -------------------------- |
| **Qwen 2.5 1.5B** | 8GB+         | Fastest | Good      | Quick tasks, 8GB Macs      |
| **Gemma 3n 2B**   | 8GB+         | Fast    | Excellent | Recommended for most users |
| **Phi-3 Mini**    | 12GB+        | Medium  | Very Good | Balanced quality and speed |
| **Gemma 3n 4B**   | 16GB+        | Slower  | Very Good | Higher quality, 16GB Macs  |
| **Qwen 3 4B**     | 16GB+        | Slower  | Good      | Alternative larger model   |

## Next Steps

<CardGroup cols={2}>
  <Card title="WhisperKit (Local)" icon="microphone" href="/providers/transcription/whisperkit-local">
    Combine with WhisperKit for 100% private, offline dictation.
  </Card>

  <Card title="Create Profiles" icon="layer-group" href="/profiles/creating-profiles">
    Different providers for different scenarios.
  </Card>

  <Card title="Privacy-First Setup" icon="shield-check" href="/guides/privacy-first-setup">
    Complete guide to local-only Stenox configuration.
  </Card>

  <Card title="Cloud Enhancement" icon="cloud" href="/providers/llm/google-gemini">
    Try cloud enhancement for comparison.
  </Card>
</CardGroup>

<Note>
  **Perfect for privacy:** MLX + WhisperKit = No cloud providers, no API keys, no internet required, complete privacy.
</Note>