Tether AI Research has released BotPsy, a 0.8-billion-parameter language model tuned for function calling and agent workflows on local devices. It builds on Qwen3.5-0.8B and is intended for routine tasks involving search, files, calendars, messaging, device control and similar tools.
The team reports that BotPsy roughly doubled several agent benchmark scores over its base model. Its BFCL v4 score rose from 23.4 to 44.0, ACEBench from 30.1 to 58.7 and the reported average on selected tau-bench tasks from 23.2 to 51.5. It also led the models tested in its size class across six tool-use benchmarks. Weights are available in Safetensors and several GGUF formats for local use.
Those results come from the model’s creators and do not establish performance in every real application. The release explicitly says BotPsy still makes mistakes, is best suited to low- and medium-difficulty work and should not be used for high-stakes decisions. Its practical appeal is therefore privacy and low resource use, not a replacement for larger models on complex reasoning.