Skip to main content

​
Getting started

This guide will help you set up your development environment and run your first local AI application with Liquid Foundation Models (LFMs).

​
What you need

Before you begin, ensure your system meets these requirements:

​
System requirements

  • Operating System: Linux, macOS, or Windows (WSL recommended)
  • RAM: Minimum 8GB (16GB or more recommended for larger models)
  • Storage: At least 5GB free space for models and dependencies
  • Python: Version 3.8 or higher
Model size determines memory requirements. Smaller models like LFM2-350M can run on 4GB RAM, while larger models like LFM2-24B require 32GB or more.

​
Common tools

Depending on which examples you want to run, you may need:
  • llama.cpp: For efficient CPU and GPU inference
  • uv: Fast Python package installer and environment manager
  • Git: For cloning the repository
  • Docker: For containerized examples (optional)

​
Clone the repository

1

Clone the LFM Cookbook

Open your terminal and clone the repository:
git clone https://github.com/Liquid4All/lfm-cookbook.git
cd lfm-cookbook
2

Explore the structure

The repository is organized into main sections:
lfm-cookbook/
├── examples/          # Ready-to-run applications
├── finetuning/        # Fine-tuning notebooks and scripts
└── README.md          # Main documentation

​
Install common dependencies

Many examples use uv for fast dependency management. Install it using: 
curl -LsSf https://astral.sh/uv/install.sh | sh
For llama.cpp-based examples, you’ll need to build or install llama.cpp. See the llama.cpp documentation for platform-specific instructions.
Each example includes its own README.md with specific setup instructions. Always check the example’s README before starting.

​
Run your first example

Let’s run the Audio Transcription CLI to see LFMs in action:
1

Navigate to the example

cd examples/audio-transcription-cli
2

Install dependencies

uv pip install -r requirements.txt
3

Download the model

Follow the instructions in the example’s README to download the LFM2-Audio-1.5B model in GGUF format.
4

Run the application

python main.py --model path/to/model.gguf
Speak into your microphone and watch real-time transcription appear in your terminal!
Model download sizes vary from hundreds of MB to several GB. Ensure you have a stable internet connection and sufficient storage.

​
Explore more examples

Now that you’ve run your first example, explore more use cases:

Invoice parser

Extract structured data from invoices using vision models

Flight search assistant

Build an AI agent with tool calling capabilities

WebGPU demos

Run models entirely in your browser without installation

Mobile apps

Deploy models on iOS and Android devices

​
Understanding model formats

LFMs are available in different formats depending on your deployment target:
  • GGUF: Optimized format for llama.cpp, ideal for CPU/GPU inference
  • ONNX: Cross-platform format for mobile and web deployment
  • Hugging Face: Standard format for Python-based inference and fine-tuning
You can find all models on Hugging Face.

​
Common workflows

​
For local desktop applications

  1. Choose an example from the examples/ directory
  2. Follow the example’s README for setup
  3. Download the appropriate GGUF model
  4. Run the application with your local model

​
For mobile deployment

  1. Set up the LEAP Edge SDK
  2. Choose an Android (Kotlin) or iOS (Swift) example
  3. Follow the platform-specific setup instructions
  4. Build and deploy to your device

​
For fine-tuning

  1. Navigate to finetuning/notebooks/
  2. Open a notebook in Google Colab
  3. Prepare your dataset in the required format
  4. Run the notebook to fine-tune your model
  5. Export and deploy your custom model

​
Next steps

1

Explore local AI apps

Check out the Local AI Apps section for production-ready examples with detailed documentation.
2

Learn about mobile deployment

Dive into Mobile Deployment to build native iOS and Android applications.
3

Customize models

Visit the Fine-Tuning section to learn how to train models on your own data.
4

Join the community

Connect with other developers on Discord to share projects and get help.

​
Get help

If you encounter issues:
  • Check the example’s README for troubleshooting tips
  • Search existing issues on GitHub
  • Ask questions in the #help channel on Discord
  • Review the Liquid AI Documentation for API references
Join the #live-events channel on Discord to participate in technical deep dives and hands-on sessions with the Liquid AI team!

Build docs developers (and LLMs) love

Get started for freeTalk to us