Features

• Fast and Lightweight: Designed for CPU inference on phones, laptops, and single-board computers, with a target memory footprint below 100 MB.

• Minimal Dependencies: Built with Python and NumPy, without requiring PyTorch, CUDA, GGUF, or C++ build tools.

• 4-Expert MoE Architecture: Implements a sparse Mixture-of-Experts neural core with RoPE, SwiGLU activations, and RMSNorm.

• Offline Privacy: Supports local execution without cloud inference, telemetry, or sending user data to external servers.

• Live Internet Access: Supports real-time web search and online information retrieval when an internet connection is available.

• Android and Termux Support: Designed to run on Android devices through APK integration or Termux.

• Dual User Interfaces: Provides an interactive terminal interface and a local web-based chat interface.

• Single-File Portability: Distributed as a self-contained ".pyz" executable, with support for standalone APK deployment.

• Anti-Repetition Decoding: Includes top-p/top-k sampling, temperature controls, and 3-gram repetition masking.

• Knowledge Boundaries: Designed to acknowledge when relevant information is unavailable rather than generate unsupported answers.

• Developer-Friendly Tools: Offers CLI arguments, interactive launchers, and tools for inspecting the local knowledge base.

• Open Source: Released under the MIT License, with model weights included.