On-Device AI Engine Specifications
A complete technical reference for combining lightweight, task-specific models into a unified, offline-first personal assistant on Android.
Model Technical Cards
Filter by domain or search by model name to inspect parameter counts, memory consumption, and runtime configurations.
Master Technical Matrix
Side-by-side comparison of parameter sizes, on-disk footprints, active memory consumption, and runtime engines.
| Model | Role | Parameters | Quantization | Disk Size | Peak RAM | Runtime Engine | Hardware Target |
|---|
Interactive 8 GB RAM Allocation Simulator
Select which models you plan to run simultaneously to test whether your setup remains within safe operating limits or triggers Android's Low Memory Killer.
Total RAM Consumption
End-to-End Orchestrated Workflows
Execution flows showing data movement across individual specialized models.
Single App Architecture Blueprint
How the native Android orchestration application is constructed using Kotlin and the Android NDK.
Sherpa-ONNX (JNI)
Direct C++ bindings handling audio hardware I/O:
- Continuous Microphone Capture: 16 kHz mono buffer streaming.
- Silero VAD: Real-time zero-copy chunk evaluation.
- Whisper Tiny & IndicConformer: Fast CPU-based speech decoding.
- Piper & Kokoro: Text-to-speech output streamed directly to Android AudioTrack.
llama.cpp (NDK + Vulkan)
Compiled natively as libllama.so to run GGUF quantized models:
- Mali-G720 GPU Acceleration: Vulkan backend offloading.
- Qwen3-1.7B Orchestration: Generates strict JSON schema tool calls.
- Gemma 3 270M: Ultra-fast text classification and intent triage.
- Moondream 2 VLM: Multimodal visual reasoning and Q&A.
ONNX Runtime Mobile
Cross-platform inference engine for vision and memory models:
- BGE-small-en: 384-dimensional vector embeddings for local RAG.
- PP-OCR Mobile: Fast text detection and line extraction.
- YOLO11n: Object detection at 60+ FPS via OpenCL/NNAPI.
- MobileCLIP-S0: Zero-shot image and text similarity matching.