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

Frontier cloud-based large language models incur high and unpredictable costs while exposing users to service disruptions and limited customization, prompting a shift toward smaller, self-hosted models. BotPsy, developed by Tether AI Research, addresses these issues with a 0.8-billion-parameter model optimized for on-device deployment and function-calling tasks. The model targets low-to-medium difficulty workflows in areas such as web research, travel planning, and device control, prioritizing privacy and local data processing over reliance on third-party providers.

BotPsy achieves state-of-the-art tool-use and agentic performance within its size class, outperforming models up to 1.2 billion parameters on standardized benchmarks. Its training pipeline combines supervised fine-tuning, multi-teacher distillation, and reinforcement learning, using rule-based rewards to refine tool-calling accuracy. The model demonstrates strong instruction-following capabilities but is not intended for high-stakes applications, as noted in its intended-use documentation.

Future development will expand BotPsy’s on-device agentic capabilities, integrating with additional hardware and local services while maintaining strict privacy protocols. Planned enhancements include decentralized peer-to-peer search, local fine-tuning for user-specific workflows, and secure agent-to-agent communication within trusted networks. Offline-to-online transitions will allow task delegation to trusted nodes, ensuring user control over data and computational resources.

BotPsy’s architecture builds on Qwen3.5-0.8B and undergoes four sequential training stages: supervised fine-tuning, multi-teacher distillation, and two rounds of reinforcement learning. The model’s tool-use performance roughly doubles that of its base, with notable gains in multi-turn interactions and instruction adherence across multiple languages. While excelling in agentic tasks, it remains constrained by parameter count, particularly in general knowledge benchmarks where larger models retain an advantage.