BotPsy releases an 0.8B-parameter tool-using assistant for on-device workflows

What is changing?

Tether AI Research announced the 0.8-billion-parameter BotPsy on November 11 and published both model weights and GGUF builds on Hugging Face. Built on Qwen3.5-0.8B, the model is designed for personal-assistant workflows that call functions across search, travel, files, calendars, messaging, and device control.

BotPsy goes through supervised fine-tuning, on-policy distillation with three specialist teachers, and two reinforcement-learning stages. In Tether’s tests, it roughly doubled the tool-use performance of its base model and ranked first on five of the six compared agent benchmarks. These are provider-run evaluations and require independent validation across different tool schemas, hardware, and tasks.

The weights are available under Apache 2.0 in safetensors and bf16, Q8, and Q4 GGUF formats. The model was trained primarily in English and is positioned for tasks of low to medium difficulty. The provider says it should not be used for high-stakes automated decisions involving money, health, personal information, or other sensitive areas.