Liquid4All/cookbook

API     Try on Playground     Documentation

Examples, tutorials, and applications to help you build with our open-weight LFMs on laptops, mobile, and edge devices.

Contents

• Desktop Apps
• Browser Apps
• Mobile Apps (Android / iOS)
• Fine-Tuning Examples
• Guides
• Third-Party Apps Powered by LFM
• Community Projects
• Technical Deep Dives
• Contributing
• Support

🖥️ Desktop Apps

Python and CLI applications for running LFM models on your laptop or desktop machine.

| Name | Description | Link | |------|-------------|------| | Invoice Parser | Extract structured data from invoice images using LFM2-VL-3B | Code | | Audio Transcription CLI | Real-time audio-to-text transcription using LFM2.5-Audio-1.5B with llama.cpp | Code | | Flight Search Assistant | Find and book plane tickets using LFM2.5-1.2B-Thinking with tool calling | Code | | Audio Car Cockpit | Voice-controlled car cockpit demo combining LFM2.5-Audio-1.5B with LFM2-1.2B-Tool | Code | | LocalCowork | On-device AI agent for file ops, security scanning, OCR, and more, powered by LFM2-24B-A2B | Code | | Home Assistant | Local home assistant with tool calling, benchmarking, and fine-tuning pipeline using LFM2-350M and LFM2.5-1.2B | Code | | Voice Assistant for Mac | On-device voice assistant for Apple Silicon Macs using LFM2.5-Audio-1.5B and the LEAP SDK | Code | | Vision Assistant for Mac | On-device visual language model chat for Apple Silicon Macs using the LEAP SDK | Code |

🌐 Browser Apps

Zero-install applications running LFM models directly in the browser via WebGPU and ONNX Runtime Web.

| Name | Description | Link | |------|-------------|------| | Tool Calling | Run LFM2 entirely in your browser with WebGPU for in-browser tool calling | Code \| Demo | | Voice Assistant | Run LFM2.5-Audio-1.5B entirely in your browser for speech recognition, TTS, and conversation | Code \| Demo | | Live Video Captioning | Real-time video captioning with LFM2.5-VL-1.6B running in-browser using WebGPU | Code \| Demo | | Chain-of-Thought Reasoning | Run LFM2.5-1.2B-Thinking entirely in your browser with WebGPU for on-device chain-of-thought reasoning | Code \| Demo | | Hand & Voice Racer | Browser driving game controlled by hand gestures (MediaPipe) and voice commands (LFM2.5-Audio-1.5B), running fully local | Code | | Road Decider | Pixel-art survival racer where Liquid's d1 decision model dodges traffic against Jev or a human player | Code | | LEAP Voice Assistant | On-device voice assistant running in the browser via WebAssembly using the LEAP SDK | Code |

📱 Mobile Apps

Native examples for deploying LFM2 models on iOS and Android using the LEAP Edge SDK. Written for Android (Kotlin) and iOS (Swift), the goal of the Edge SDK is to make Small Language Model deployment as easy as calling a cloud LLM API endpoint.

Android

| Name | Description | Link | |------|-------------|------| | LeapChat | Chat app with real-time streaming, persistent history, and modern UI | Code | | Voice Assistant | Audio input and output with LFM2.5-Audio-1.5B for on-device AI inference | Code | | LeapKoogAgent | Integration with Koog framework for AI agent functionality | Code | | SloganApp | Single turn marketing slogan generation with Android Views | Code | | ShareAI | Website summary generator | Code | | Recipe Generator | Structured output generation with the LEAP SDK | Code | | VLM Example | Visual Language Model integration | Code |

iOS

| Name | Description | Link | |------|-------------|------| | LeapChat | Chat app with real-time streaming, conversation management, and SwiftUI | Code | | LeapSloganExample | Basic LeapSDK integration for text generation in SwiftUI | Code | | Recipe Generator | Structured output generation | Code | | Voice Assistant | Audio input and output with LFM2.5-Audio-1.5B for on-device AI inference | Code | | Vision Assistant | Visual language model chat on iOS using the LEAP SDK | Code | | Telco Triage | Edge-first home internet support assistant with on-device LFM routing, local tools, and cloud-assist handoff | Code |

🎯 Fine-Tuning Examples

We recommend fine-tuning LFM3 for your specific use case to achieve the best results. Below you'll find Colab notebooks and Python scripts for customizing LFM models with your own data.

| Name | Framework | Description | Link | |------|-----------|-------------|------| | Text Models | | | | | Supervised Fine-Tuning (SFT) for structured extraction | TRL (Docs) | Fine-tune LFM3-1.7B with LoRA to turn messy documents in many languages into JSON that follows a schema | Notebook | | Continued Pre-Training (CPT) for a new language | TRL (Docs) | Continue pre-training LFM3-1.7B with LoRA on Turkish Wikipedia, a language it doesn't officially support | Notebook | | Direct Preference Optimization (DPO) for grounded answers | TRL (Docs) | Teach LFM3-1.7B to answer only from the retrieved documents, and to say so when the answer isn't there | Notebook | | Group Relative Policy Optimization (GRPO) for tool calling | TRL (Docs) | Train LFM3-1.7B as a phone assistant that picks the right calls among 108 tools, with a verifiable reward and LoRA | Notebook | | Continued Pre-Training (CPT) for text completion | Unsloth (Docs) | Teach models domain-specific knowledge and creative writing styles | Notebook | | Continued Pre-Training (CPT) for translation | Unsloth (Docs) | Adapt models to specific languages or translation domains using domain data | Notebook | | Supervised Fine-Tuning (SFT) | Unsloth (Docs) | Memory-efficient SFT using Unsloth with LoRA for 2x faster training | Notebook | | Supervised Fine-Tuning (SFT) | TRL (Docs) | Supervised fine-tuning using Hugging Face TRL library with parameter-efficient LoRA | Notebook | | Supervised Fine-Tuning (SFT) for Mixture-of-Experts | Halo (Docs) | Supervised fine-tuning of LFM2.5-8B-A1B with two-way expert parallelism using Halo | Notebook | | Direct Preference Optimization (DPO) | TRL (Docs) | Align models with human preferences using direct preference optimization with LoRA | Notebook | | Antidoom Training (FTPO) | TRL (Docs) | Eliminate doom loops with final token preference optimization (FTPO) using the Antidoom pipeline | Notebook | | Group Relative Policy Optimization (GRPO) | TRL (Docs) | Train reasoning models using Group Relative Policy Optimization with rule-based rewards | Notebook | | Group Relative Policy Optimization (GRPO) | TRL (Docs) | Train a 350M model for better structured output compliance | Notebook | | Group Relative Policy Optimization (GRPO) | Unsloth (Docs) | Train reasoning models using Group Relative Policy Optimization for verifiable tasks | Notebook | | Decision Models | | | | | Supervised Fine-Tuning (SFT) | TRL (Docs) | Learn how to fine-tune a decision model (d1-3B) on your own labels with LoRA, using intent classification on Banking77 as the example | Notebook | | Supervised Fine-Tuning (SFT) | Unsloth (Docs) | Learn how to fine-tune a decision model (d1-3B) on your own labels with Unsloth and LoRA, using intent classification on Banking77 as the example | Notebook | | Vision Models | | | | | Supervised Fine-Tuning (SFT) for receipt extraction | TRL (Docs) | Fine-tune LFM3-VL-2B with LoRA to read photos of receipts into JSON that follows a schema | Notebook | | Group Relative Policy Optimization (GRPO) for formula transcription | TRL (Docs) | Train LFM3-VL-2B with LoRA to transcribe pictures of formulas into LaTeX, with a format reward and a LaTeX similarity reward | Notebook | | Supervised Fine-Tuning (SFT) | Unsloth (Docs) | Supervised fine-tuning for LFM2-VL models on custom image-text datasets | Notebook | | Supervised Fine-Tuning (SFT) | TRL (Docs) | Supervised fine-tuning for LFM2.5-VL models using TRL with parameter-efficient LoRA | Notebook | | Supervised Fine-Tuning (SFT) | Halo (Docs) | Supervised fine-tuning for LFM2.5-VL-3B on CPPE-5 for object detection | Notebook | | Wind Turbine Blade Inspection | TRL (Docs) | Fine-tune LFM3-VL-2B with SFT and GRPO to find and classify surface defects on wind turbine blade photos, and compare it with frontier models | Code | | Satellite VLM Fine-Tuning | leap-finetune | Fine-tune LFM2.5-VL-450M on satellite imagery for VQA, grounding, and captioning using Modal | Code | | Wildfire Prevention | leap-finetune | Build a wildfire risk detection system using LFM2.5-VL-450M and Sentinel-2 satellite imagery, with fine-tuning and on-device inference | Code | | Audio Models | | | | | Supervised Fine-Tuning (SFT) for voice commands | HF Transformers | Fine-tune LFM3-Audio-2B with LoRA to map spoken smart-home commands directly to function calls | Notebook | | Supervised Fine-Tuning (SFT) for a new voice | HF Transformers | Teach LFM3-Audio-2B a new, consistent voice by training its speech output on LJSpeech | Notebook | | LFM2.5-Audio Fine-Tuning | liquid-audio | Fine-tune LFM2.5-Audio-1.5B on the OHF-Voice dataset to map speech directly to Home Assistant function calls, with on-device GGUF inference | Code | | Omni Models | | | | | Supervised Fine-Tuning (SFT) for voice commands | HF Transformers | Fine-tune LFM3-Omni-2B with LoRA to map spoken smart-home commands directly to function calls | Notebook | | Supervised Fine-Tuning (SFT) for spoken image Q&A | HF Transformers | Fine-tune LFM3-Omni-2B to answer spoken questions about medical images | Notebook | | Encoder Models | | | | | Document Classification | HF Transformers | Fine-tune LFM2.5-Encoder for custom long-document multi-label classification | Code |

📓 Guides

Notebooks walking through the capabilities of specific LFM models with Hugging Face Transformers.

| Name | Description | Link | |------|-------------|------| | Vision Capabilities | Image description, visual question answering, multi-image comparison, text reading, object localization, and tool calling with LFM2.5-VL | Notebook | | Longevity Models | Interpret clinical, DNA methylation, transcriptomic, proteomic, and genetic aging data with LFM2-1.2B-Longevity and LFM2-2.6B-Longevity | Notebook |

Third-Party Apps Powered by LFM

Production and open-source applications that support LFM models as an inference backend, among other providers.

| Name | Description | Link | |------|-------------|------| | DeepCamera | Open-source AI camera system for local vision intelligence with facial recognition, person re-ID, and edge deployment on Jetson and Raspberry Pi | Code | | Osaurus | Native macOS AI harness for managing agents, memory, tools, and identity locally, with support for LFM models via MLX on Apple Silicon | Code | | SelfLink | Native iOS privacy-first journaling app | Website |

🌟 Community Projects

Open-source projects built by the community showcasing LFMs with real use cases.

Fine-tuning

| Name | Description | Link | |------|-------------|------| | LFM2.5 Mobile Actions | LoRA fine-tuned LFM2.5-1.2B that translates natural language into Android OS function calls for on-device mobile action recognition | Code | | Food Images Fine-tuning | Fine-tune LFM models on food image datasets | Code | | LFM2-KoEn-Tuning | Fine-tuned LFM2 1.2B for Korean-English translation | Code | | SFT + DPO Fine-tuning | Teaching a 1.2B model to be a grumpy Italian chef: SFT + DPO fine-tuning with Unsloth | Code | | LFM2-2.6B Mr. Tic Tac Toe | Fine-tune LFM2-2.6B with reinforcement learning to play tic-tac-toe | Code |

Deployment

| Name | Description | Link | |------|-------------|------| | LF