Tether AI Research has introduced BotPsy, a 0.8B-parameter language model aimed at function calling and agent workflows on local devices. It is post-trained from Qwen3.5-0.8B. The team positions it as a personal assistant that uses tools for web research, email and messaging, calendar, files, commerce and finance, and what it calls business and customer operations.
What the benchmarks say
All figures below are the authors' own evaluations, run in thinking mode, with no independent check in the source.
- Against its base model: BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and the τ-bench average from 23.2 to 51.5.
- Against a larger model: BotPsy tops Qwen3.5-2B-Think, which is 2.5 times its size, on five of six agent benchmarks, including ToolSandbox (64.6 vs 41.2). On Claw-Eval the larger model stays ahead, 48.9 to 46.5.
- Multi-turn work: On τ²-bench telecom, where the agent guides a user through troubleshooting, it scores 82.5 against 69.3 for the 2B model.
- Knowing when to refuse: It scores 86.Ranked 1, with the top result among the models tested on the BFCL v4 hallucination assessments, which verify that a model refuses to act when no suitable tool is available.
- Reasoning cost: On GPQA Diamond it needs 3,690 reasoning tokens per correct answer, against 11,563 for its base. The authors expect this to cut on-device latency.
Safety numbers
In automated red-teaming across six harm categories, BotPsy gave unsafe responses 21% of the time, against 36% for its base. Under prompt injection, the case that matters for agents reading untrusted tool output, the rate fell from 46% to 31%. Those rates are lower than the base model's but far from zero.
Limits the vendor states
The team says the model suits low to medium difficulty tasks and makes mistakes. It must not be relied upon for automated decisions where safety is at risk or the stakes are high, nor for uses that touch on finances, private data, or medical matters. Multi-step tasks score lower than single tool calls, and open-ended deep research is weak: BrowseComp-Plus is 4.5 out of 100. English is the recommended language. On Multi-IF, Hindi scores 24.5 and Russian 57.6, well below the 72-80 range in English and several other languages. General knowledge still tracks model size, so the model is meant to call tools for facts it lacks.
Availability
The post says BotPsy will be released with open weights under Apache 2.0 for research and educational use, in bf16, Q8 and Q4 precisions. Formats are Hugging Face safetensors and GGUF files for the QVAC SDK and llama.cpp. The research-and-educational wording is worth reading closely before any commercial deployment. The source gives no price and no hosted API.