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

# Choosing an LLM

> Run local LLMs for ESP32 AI toys with MLX: Ministral, Gemma, Qwen, Llama, gpt-oss, Muse Glimmer and any model from mlx-community.

The brain of your toy is any LLM from the [`mlx-community`](https://huggingface.co/mlx-community) organization on Hugging Face. OpenToys loads it with `mlx-lm`, and falls back to `mlx-vlm` for models that need it.

| | |
| - | - |
| **Default model** | `mlx-community/Ministral-3-3B-Instruct-2512-4bit` |
| **Also works with** | Gemma 4, Qwen3, Llama, gpt-oss, Muse Glimmer and other `mlx-community` models |
| **Context window** | 4096 tokens |

## Models

> **[Ministral 3](https://elatoai.com/docs/local/llm/ministral.md)**
>
> Default. Small and fast.

> **[Gemma](https://elatoai.com/docs/local/llm/gemma.md)**
>
> Google's open models.

> **[Qwen3](https://elatoai.com/docs/local/llm/qwen3.md)**
>
> Multilingual from Alibaba.

> **[Llama](https://elatoai.com/docs/local/llm/llama.md)**
>
> Meta's open models.

> **[gpt-oss](https://elatoai.com/docs/local/llm/gpt-oss.md)**
>
> OpenAI's open-weight model.

> **[Muse Glimmer](https://elatoai.com/docs/local/llm/muse-glimmer.md)**
>
> Large 30B model.

## Switching models

Open the **Models** page in OpenToys, pick an `mlx-community` repo and the app hot-swaps it without restarting your session.

> [!TIP]
>
> Smaller 4-bit models respond faster, which matters for a toy. Start with a 3B model and size up if your Mac has memory to spare.

## Personalities

Each character is a system prompt plus a voice. The prompt is what makes a toy a storyteller, a game host or a tutor. Default characters live in [`personalities.json`](https://github.com/akdeb/OpenToys/tree/main/app/src/assets).


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