BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

Frontier cloud language models come with high operational costs and potential risks due to provider modifications, denials, or deprecations. These models also limit customization, raise privacy concerns, and demand significant computational power for many routine tasks. In response, Tether AI Research has introduced BotPsy, a compact 0.8 billion parameter model designed for on-device deployment.

BotPsy is tailored for function calling and agentic workflows, excelling in tool-use and agentic performance for its size, surpassing models up to 1.2 billion parameters. It functions as a personal assistant for tasks ranging from web search to email management, prioritizing privacy and efficiency. Built upon the Qwen3.5-0.8B model, BotPsy undergoes four stages of training, including supervised fine-tuning, multi-teacher on-policy distillation, and reinforcement learning.

This iterative process aims to enhance its capabilities while maintaining user data privacy. Future developments by Tether AI Research will focus on expanding on-device agentic functionalities and integrating with various hardware and local services to create a more robust, decentralized personal AI ecosystem.