[Draft] Deprecate LEAP SDK and add native llama.cpp developer guides

[tuliren]

Summary

The LEAP SDK is being deprecated. This PR de-emphasizes it in the docs (kept, but archived) and replaces the SDK developer guides with guides that use llama.cpp directly.

1. New "Build with llama.cpp" section (deployment/on-device/llama-cpp/)

Page	Replaces (LEAP SDK)
iOS & Android — XCFramework / NDK integration, C API load → sample → stream loop in Swift, upstream Kotlin binding	Quick Start, Model Loading
Desktop & Server Apps — llama-server sidecar (Node/Python/.NET), in-process bindings (llama-cpp-python, node-llama-cpp, LLamaSharp), hybrid local + cloud routing	Desktop & Native Platforms, OpenAI-Compatible Client
Chat & Streaming — OpenAI-compatible API, sampling table per model family, prompt caching, native multi-turn	Conversation & Generation, Cloud AI Comparison
Function Calling & Agents — --jinja tool calling (llama.cpp parses LFM2/LFM2.5 tool-call formats natively in common/chat.cpp), agent loop	Function Calling, AI Agent Usage Guide
Structured Output — JSON schema / GBNF, llama_sampler_init_grammar	Constrained Generation
Vision & Audio — mmproj vision via server/CLI/mtmd, LFM2.5-Audio runners	Messages & Content, Voice Assistant
Migrating from LEAP SDK — concept-by-concept mapping table	—

The llama.cpp overview page links to these with a card group.

2. LEAP SDK deprecated in place

• "Leap SDK" tab removed; Model Bundling moved into the LFM tab (after Fine-tuning).
• All deployment/on-device/sdk/* pages and the changelog stay on disk (URLs keep resolving per the link contract) with a <Warning> banner and noindex: true.
• SDK overview rewritten as a short "LEAP SDK (archived)" notice.
• New "Deprecated SDKs" section on the Deprecations page.

3. Cross-references updated

FAQs, hardware evaluation, migration guide, model library, connect-AI-tools snippet, README, quickstart-widget notes. Android example pages keep their content but get a "built with the deprecated LEAP SDK" note.

Verification

• npm run snapshot:check passes (7 new URLs added; nothing removed).
• mintlify broken-links: only 2 pre-existing failures in untouched files (CLAUDE.md relative link, a model-bundling image).
• All new/changed pages render (HTTP 200, correct titles) on mintlify dev.
• llama.cpp API details (C function signatures, common/chat.cpp LFM2 detection, release assets, HF GGUF filenames) were verified against upstream master / the HF API on 2026-08-27.

Open questions for reviewers

1. Model Bundling (leap-bundle) — kept as-is and moved into the LFM tab. Is it also being deprecated with the LEAP platform?
2. LEAP Model Library links (leap.liquid.ai/models) on the model overview pages were left untouched. Remove if the library is going away.
3. Android examples — kept in the Examples tab with a deprecation note. Happy to drop them from navigation instead.
4. repeat_penalty vs repetition_penalty — llama-server only reads repeat_penalty (tools/server/server-schema.cpp); repetition_penalty is silently ignored. The new pages use repeat_penalty. CLAUDE.md and the existing llama-cpp.mdx say repetition_penalty for llama-server — I left those alone but they should probably be fixed in a follow-up.
5. The existing llama-cpp.mdx uses lowercase GGUF filenames (lfm2.5-1.2b-instruct-q4_k_m.gguf); the files on HF are LFM2.5-1.2B-Instruct-Q4_K_M.gguf (paths are case-sensitive). Not touched here.

🤖 Generated with Claude Code

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
liquid-docs	🟢 Ready	View Preview	Aug 27, 2026, 10:13 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.