BotPsy tests compact agents for private local assistants
Tether AI Research released an open 0.8B tool-using model designed for on-device assistants, reporting strong benchmark gains but limited scope.
What happened
Tether AI Research introduced BotPsy, a 0.8-billion-parameter language model built on Qwen3.5-0.8B for function calling and agent workflows. The model is designed to run locally rather than depend on a cloud endpoint, with intended uses including web search, calendars, files, messaging, commerce, device control and other bounded assistant tasks.
The release uses a four-stage post-training recipe: supervised fine-tuning, multi-teacher on-policy distillation, and two reinforcement-learning stages. The teachers were specialized for tool use, reasoning and instruction following. Tether says BotPsy outperformed tested models of up to 1.2 billion parameters across six tool-use and agent benchmarks, and exceeded Qwen3.5-2B-Think on five of six tests. Against its base model, the largest reported gains were on BFCL v4, ACEBench and τ-bench.
The model is released under Apache 2.0 in bf16, Q8 and Q4 formats, with Hugging Face safetensors and GGUF files for the QVAC SDK and llama.cpp. Tether also reports a lower unsafe-response rate than the base model in automated red-teaming, including under prompt-injection tests, but the figures are internally produced rather than independently replicated.
Why it matters
BotPsy is significant because it treats tool execution—not general conversation—as the optimization target for a sub-billion model. That could make local assistants viable on phones, PCs and embedded systems where privacy, offline operation and predictable hardware costs matter. The release does not show that a tiny model is ready to manage unrestricted personal workflows: its authors limit it to low- and medium-difficulty tasks, and tool errors remain consequential. The practical question is whether its benchmark advantage survives real APIs, languages, permissions and adversarial inputs.
Uncle Cat take
BotPsy’s important number is not 0.8B; it is the reported 21% unsafe-response rate, which still leaves local tool access far from automatic trust.