Manthan
Manthan (Sanskrit: मंथन, “churning”) is a private, on-device AI assistant. Chat with local LLMs, reason over images, talk to your own documents, and dictate with your voice — 100% on-device. Switch on airplane mode and everything still works.
Highlights
- No network during inference — no API keys, accounts, or telemetry
- Pluggable engines: LiteRT-LM/MediaPipe and llama.cpp behind one interface
- On-device RAG with chunking, embeddings, ObjectBox HNSW, and citations
- Multimodal image input and on-device speech-to-text
- Model manager with resumable downloads and checksum verification
Why Manthan
Most AI apps are thin clients that stream your data to someone else’s servers. Manthan is the opposite — every token is generated on the device in your hand. The architecture prioritizes offline-first persistence, model switching, conversational UX, and device-friendly performance.
Inference & RAG
Inference is pluggable via LiteRT/MediaPipe Gemma models and llama.cpp/GGUF through Dart FFI. Documents are chunked and embedded locally, indexed with ObjectBox HNSW vector search, and answered with grounded responses and citations — all without leaving the device.