Tether AI Research presents BotPsy, a 0.8B model for on device agents

Tether AI Research presented BotPsy, a model with 0.8 billion parameters for simple tasks such as web search, calendar and email. It operates on a device and calls software tools, and Tether plans to release the weights. Tether gives scores that beat each tested model of 1.2 billion parameters or less on six tool use tests.

How it works

BotPsy is built on Qwen3.5-0.8B. It is a personal assistant that uses tools, and it suits tasks from small to moderate. Tether plans to release the weights for research and education.

The numbers

• ACEBench: 30.1 to 58.7.
• Dangerous replies in red team tests: 36% to 21%.
• It beat a larger model of the same family on most tests.

What is not known

All figures are from Tether. The model makes mistakes. Tether tells that it is not for high stakes use.

Insight: Small on device agents are a race for the local stack

Google gives developers a C++ layer for on device models. Tether adds a small model and plans its own search and tools to adapt models to a user. This reads as a race to own the on device stack, where the maker of the layer gets the users. Builders get open weights, but must monitor which tool chain locks them in.