LEAP is an on-device AI development platform from Liquid AI, an MIT-born startup. It covers the whole path for the company’s lightweight foundation models — the LFM (Liquid Foundation Models) family — from model discovery and fine-tuning to bundling and app integration. Its defining trait is that you can call a model sitting on the device from just a few lines of code, much as you would call a cloud API. Targets include iOS and Android apps as well as laptops and small devices in the Raspberry Pi class. It launched as an alpha in July 2025 and has since broadened its coverage to vision, audio, and tool calling alongside the LFM2.5 generation of models.
Key Features
- Model library and model search: Browse and download model bundles already optimized for LEAP. A “Best Model Search” helps you find a model that fits both your task and your device constraints (memory, speed)
- Leap Edge SDK for app integration: An SDK that loads and queries on-device models as easily as calling a cloud API. Built on Kotlin Multiplatform, it exposes the same API across iOS, macOS, Android, JVM desktop, Linux, and Windows, with experimental WebAssembly support
- First-class support for the LFM2 / LFM2.5 family: Because the team that trains the models also ships the SDK, chat templates and special-token handling are correct out of the box. Beyond text, there are thinking, Japanese-focused, vision-language, and audio variants
- Fine-tuning and model bundling: Adapt LFM2 and other LEAP models to your use case with GPU-optimized scripts, then package the result into a distributable bundle
- Tool calling and structured output: Function calling is supported, and constrained generation enforces JSON schemas at decode time through Kotlin annotations and Swift macros. This makes agent-style processing that stays entirely on the device easier to build
- On-device testing with Liquid Apollo: A native app lets you try models on a real handset so you can check response speed and quality before integrating
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Free | $0 (no cost, ever) | Model search, downloads from the model library, fine-tuning tools, model bundling, LEAP Edge SDK |
| Enterprise | Contact sales | Everything in Free, plus dedicated support for models that need proprietary data or specialized expertise, and deployment assistance for complex environments |
Pricing is as of August 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- Core features are free, so even solo developers can try on-device AI without cost
- Inference happens on the device, avoiding bandwidth costs, per-token API billing, and network latency
- Input data never leaves the device, which suits privacy-sensitive use cases
- Works offline, so features keep running where there is no signal or on a plane
- Because the model vendor also ships the SDK, mismatches between model and runtime are less likely
⚠️ Cons
- The available models center on Liquid AI’s own LFM family, so it is not the tool for freely running third-party large models
- Being lightweight models, they have a ceiling in knowledge and reasoning compared with large cloud models
- Device differences (memory, chip generation) change which models are usable and how fast they feel, adding testing overhead
- Bundling a model with your app increases binary size, which requires some distribution planning
- As a relatively new platform, implementation examples and write-ups are still limited
Comparison with Similar Services
| Criterion | LEAP | Google AI Edge (LiteRT) | Apple Foundation Models | Ollama |
|---|---|---|---|---|
| Provider | Liquid AI | Apple | Ollama | |
| Main targets | iOS / Android / PC / edge devices | Android-first (iOS also) | Apple devices only | PCs and servers |
| Main models | LFM2 / LFM2.5 family | Gemma and others | OS-bundled on-device models | Open models broadly |
| Fine-tuning | Official scripts provided | Supported (separate tooling) | Limited (adapter-based) | Via external tools |
| Mobile app integration | Strong (SDK-first) | Strong | Strong (Apple only) | Not intended |
| Price | Free (Enterprise on request) | Free | Free (part of the OS) | Free |
Who Is It For
- Mobile app developers who need AI features to stay on the device because of connectivity or privacy constraints
- Solo developers and small teams that want AI features without ongoing per-usage API bills
- Edge developers embedding AI into small devices in the Raspberry Pi class or into industrial equipment
- Developers who want the same AI feature on both iOS and Android with as much shared implementation as possible
- Teams that want fine-tuning through distribution of lightweight models handled in one flow
Summary
LEAP’s strength is its structure: the company that builds the lightweight foundation models also takes care of how you distribute and embed them. Model discovery, fine-tuning, bundling, and SDK integration sit on a single path, which flattens the usual on-device AI stumbling block of model conversion and runtime compatibility. It is not a replacement for large cloud models, but if privacy, offline operation, or cost is part of your requirements, the free tier’s model library and Edge SDK are worth trying first.