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

Why it matters

Tether AI Research has released BotPsy, a compact agentic language model built on Qwen3.5-0.8B and intended to run on-device rather than in the cloud. The model is positioned as a personal assistant that handles routine tool-based tasks such as web search, calendar management, messaging, commerce, and device control. The team frames the work as a response to the costs, privacy risks, and limited customization that come with relying on proprietary frontier models, arguing that many everyday tasks do not require frontier-level intelligence and can instead be handled locally by small, specialized agents. BotPsy roughly doubles the agentic performance of its base model across several tool-use benchmarks and outperforms other small models in its weight class, including some larger than itself, on nearly all tests reported. It also shows stronger instruction following, greater resistance to hallucinating tool calls, and improved safety in automated red-teaming compared with its base model. The release is intended for research and educational use, with open weights under an Apache 2.0 license, and the team cautions that it is not suited to high-stakes applications.