Migrating from LEAP SDK

The LEAP SDK is deprecated and no longer receives new releases. It was a Kotlin Multiplatform wrapper around llama.cpp; everything it did is available directly from llama.cpp, which supports LFM2 / LFM2.5 models, LFM2-VL projectors, and LFM2 tool-call parsing upstream. Existing LEAP SDK artifacts remain on Maven Central and GitHub, and the archived reference is at LEAP SDK (archived).

Concept mapping

LEAP SDK	llama.cpp equivalent	Guide
`LeapModelDownloader.loadModel(modelName:, quantizationType:)`, LEAP Model Library bundles	Download the GGUF from Hugging Face (`LiquidAI/<model>-GGUF`) with your platform’s downloader; `llama-server -hf <repo>:<quant>` on desktop	iOS & Android
`ModelRunner`	`llama_model` + `llama_context` (C API) or a running `llama-server`	iOS & Android, Desktop & Server Apps
`Conversation`, `ChatMessage`	An OpenAI-style `messages` array; the chat template is applied by `llama-server --jinja` or by your code	Chat & Streaming
`conversation.generateResponse(...)` streaming `MessageResponse.Chunk`	`stream: true` on `/v1/chat/completions`, or the `llama_decode` / `llama_sampler_sample` loop	Chat & Streaming
`GenerationOptions` (`temperature`, `topK`, `minP`, `repetitionPenalty`, `maxTokens`)	Request fields `temperature`, `top_k`, `min_p`, `repeat_penalty`, `max_tokens`; `llama_sampler_init_*` in the C API	Sampling parameters
Per-checkpoint sampler defaults from the bundle manifest	Values on each Hugging Face model card (summarized in the sampling table)	Sampling parameters
`LeapFunction`, `MessageResponse.FunctionCalls`, `LFMFunctionCallParser` / `HermesFunctionCallParser`	OpenAI `tools` / `tool_calls` with `llama-server --jinja` (LFM2 and LFM2.5 formats parsed natively); `common_chat_parse()` in-process	Function Calling & Agents
`@Generatable` / `@Guide` constrained generation, `jsonSchema` in `GenerationOptions`	`response_format: {"type": "json_schema", ...}` or a GBNF `grammar`; `llama_sampler_init_grammar()` in the C API	Structured Output
`ChatMessageContent.Image` (JPEG bytes)	`image_url` content part with a base64 `data:` URI; `mtmd` API in-process, with the model’s `mmproj-*.gguf`	Vision & Audio
`ChatMessageContent.Audio` (WAV) and `MessageResponse.AudioSample`	Liquid’s `llama-liquid-audio-cli` / `llama-liquid-audio-server` runners for LFM2.5-Audio	Vision & Audio
`CacheOptions` / KV cache reuse	`cache_prompt` (default on) and `--cache-reuse` in `llama-server`; keep the context alive between turns in the C API	Multi-turn conversations
`ModelLoadingOptions` (`nCtx`, `nThreads`, `nGpuLayers`, `useMmap`)	`llama_context_params.n_ctx` / `n_threads`, `llama_model_params.n_gpu_layers` / `use_mmap`; `-c`, `-t`, `-ngl` flags	Tune for mobile
`leap-openai-client` (hybrid on-device + cloud)	Any OpenAI client pointed at `llama-server` locally and a hosted endpoint remotely	Hybrid routing
`leap-ui` voice assistant widget	No drop-in replacement. Pair the audio runner with your platform’s recording/playback APIs.	Vision & Audio
Desktop targets (JVM, Kotlin/Native, Windows, Linux)	`llama-server` sidecar or a binding for your language (Python, Node.js, .NET, Rust, Go, Java)	Desktop & Server Apps

Platform notes

iOS / macOS
Android
Desktop / server

Replace the LeapSDK Swift package with llama.xcframework from a llama.cpp release and call the C API directly (import llama). Metal is enabled in the prebuilt framework. The minimum deployment target drops to iOS 16.4 / macOS 13.3. See iOS & Android.

Replace the ai.liquid.leap:* Maven dependencies with llama.cpp built through the NDK — either the upstream examples/llama.android library module or your own CMake/JNI wrapper. Model downloads move to WorkManager / DownloadManager. See iOS & Android.

Launch llama-server as a sidecar process or use an in-process binding. The OpenAI-compatible API gives you chat, streaming, tools, and JSON schema without any native code. See Desktop & Server Apps.

Models and bundles

LEAP SDK loaded “bundles” (a GGUF plus a manifest with companion files) from the LEAP Model Library. The same weights are published as plain GGUF on Hugging Face under LiquidAI: the language model file, plus mmproj-* for vision models and the encoder / vocoder / tokenizer files for audio models. If you already have downloaded bundles, the .gguf files inside them load directly in llama.cpp. Fine-tuned models continue to work: convert them with convert_hf_to_gguf.py as described in Converting Custom Models.

Archived reference

The LEAP SDK pages remain online for teams still shipping the SDK but are no longer maintained: Overview, Quick Start, Changelog. The Android example apps under Examples were built with the LEAP SDK and are kept as architectural references.

Models

Fine-tuning

Model Bundling

Edge Inference

Build with llama.cpp

GPU Inference

Cloud Inference

Guides

Help

Migrating from LEAP SDK

Concept mapping

Platform notes

Models and bundles

Archived reference

​Concept mapping

​Platform notes

​Models and bundles

​Archived reference

Concept mapping

Platform notes

Models and bundles

Archived reference