Tether AI Research has introduced BotPsy, a small model trained to call software tools and carry out multistep tasks on a device. It builds on Qwen3.5 and targets work such as calendars, files and settings. Model weights carry an Apache license, while training code and data remain outside the release and sensitive automated decisions fall outside intended use.
Tether AI Research introduced BotPsy, a language model built from Qwen3.5-0.8B for function calling and multistep tool use on a device. It targets low-to-medium difficulty tasks involving files, calendars, email and device settings. The model has 0.8 billion parameters. Its small scale is part of a plan for local personal assistants that can choose a software tool and supply the arguments it needs.[1], [2]
The training sequence begins with supervised examples covering single calls, parallel calls, several linked actions and situations requiring no tool. BotPsy then generates responses with feedback from specialist teacher models trained for tool use, reasoning and instruction following. Reinforcement learning follows in two stages: first checking tool names and arguments, then expanding to mathematics, code, structured outputs and simulated workplace tasks. The sequence combines demonstrations with rule-based and verifier-based rewards.[1]
The model weights carry the Apache 2.0 license, while training code and data are excluded from this release. English is the main training and recommended usage language. The developers identify Hindi, Russian, accumulating instructions and open-ended research as difficult areas. Sensitive automated decisions involving health, money or personal information fall outside intended use. Research and education remain described applications, with further local search and personalisation presented as development plans.[1]
