On-Device GGUF Engine
Run open-weights models (Qwen 2.5, Llama 3.2, Gemma 2, Phi-3 Mini) with zero internet connection, real-time tok/s metrics, and SAF storage persistence.
Sednium Oorty.
One Interface.
Every AI Model.
Run native quantized GGUF models directly on your phone with real-time token telemetry. Auto-save every conversation into an open Obsidian-compatible Markdown vault with YAML frontmatter, execute autonomous MCP tools, and orchestrate top cloud models seamlessly.
100%
Offline GGUF
384-D
Vector Recall
8+
Cloud AI Providers
.md
Obsidian Vault
Explain StateFlow in Kotlin Jetpack Compose
Qwen 2.5 (0.5B)
28.4 tok/s
val state by viewModel.uiState
.collectAsStateWithLifecycle()
Text("Result: ${state.data}")
⚡ 0.78s TTFT • 420MB / Safe
Vault ✓
---
id: "a1b2c3d4-e5f6"
title: "StateFlow Compose"
model: "Qwen2.5-0.5B"
tags: ["kotlin", "compose"]
---
# StateFlow Compose
## 🤖 Oorty — 17:02 | ~520 tok
StateFlow is an observable holder...$ mcp::recall_from_vault
query: "Room Database KSP"
[SUCCESS] Score 0.892
🤖 Injected 480 tokens into context.
Gemini 2.5 Flash
Google AI Studio
Claude 3.7 Sonnet
Anthropic BYOK
GPT-4o
OpenAI Direct
DeepSeek R1
Groq 400 tok/s
Ask local model anything...
🎙️
➤
Unifying On-Device & Cloud AI Ecosystems
⚡ Native GGUF (llama.cpp)
Google LiteRT
Google Gemini
Anthropic Claude
OpenAI GPT-4o
Groq LPU
NVIDIA NIM
Sednium Rosette
Core Architecture
Built for Modern AI Engineers & Note-Takers
From offline GGUF execution to Markdown databases, every layer is engineered for performance and privacy.
Obsidian Markdown Vault
Every chat is dual-written as standard .md files into Documents/Oorty/chats/ with structured YAML frontmatter, headers, tags, and speed logs.
Dynamic RAM Watchdog
Active hardware memory safety: checks real-time available RAM via ActivityManager.MemoryInfo to prevent system out-of-memory crashes.
Hybrid Semantic AI Recall
384-dimensional vector embeddings rank past conversations using Cosine similarity to automatically inject relevant memory into fresh prompt turns.
MCP Agentic Framework
Connect to Model Context Protocol (MCP) servers and execute multi-step tool calls even on compact on-device models with JSON schema injection.
Biometric App Lock & BYOK
All API keys, model presets, and chat vaults are stored locally on your device with optional hardware biometric fingerprint and face unlock.
Hardware Compatibility
RAM & Model Fit Calculator
Select your phone's memory to find the perfect on-device GGUF model that runs smoothly without crashing.
Direct Obsidian Sync
Your Chats as an Open Markdown Database
Unlike proprietary chat apps that lock your conversations into closed cloud databases, Oorty writes standard .md files straight to your device's Documents/Oorty/chats/ folder.
- Obsidian Mobile Ready: Point Obsidian to
Documents/Oorty/to search, tag, and link your AI discussions instantly. - Structured YAML Frontmatter: Every note contains title, timestamp, tokens, model, and TF-IDF generated topic tags.
- Zero Vendor Lock-In: Move, backup, sync via Git or Syncthing, or edit in any markdown editor.
📱 Obsidian Mobile Vault View
📁 Documents/Oorty/
📁 chats/
📄 kotlin-stateflow-4fa4.md
📄 room-database-setup-328a.md
📄 system-architecture-9661.md
📁 models/
FAQ
Frequently Asked Questions
How does on-device GGUF inference work in Oorty?
Oorty integrates a native llama.cpp engine via Kotlin Coroutines. You can download quantized GGUF models directly from HuggingFace within the app or choose custom files from your device storage. Generation streams token-by-token directly in RAM without sending any data to the cloud.
Where are my chats saved and can I view them in Obsidian?
Yes! Every chat session is saved in public storage at Documents/Oorty/chats/ as standard .md Markdown notes with YAML frontmatter. Open the Obsidian mobile app and tap "Open folder as vault" on Documents/Oorty/ to immediately explore all your chats.
What happens if I try to load a model too large for my phone?
Oorty's built-in HardwareChecker evaluates your device's currently available RAM. If the model requires more memory than is safely available, Oorty displays a warning dialog detailing the RAM deficit and suggests lighter 4-bit quantizations to prevent system crashes.
Do I need to pay or buy a subscription to use Oorty?
No! Oorty is 100% free and open-source under the MIT license. For local GGUF models, it runs completely free offline. For cloud models (Gemini, Claude, GPT-4o, Groq), you simply supply your own API keys (BYOK) and pay direct provider rates with zero markup.
Take Control of Your AI Intelligence
Run offline GGUF models, store your knowledge in Obsidian, and orchestrate top cloud LLMs in one unified workspace.