Most AI assistants that can act for you, booking, searching or sending messages, run on someone else's servers. Tether AI Research, part of the company best known for the Tether stablecoin, has published BotPsy, a 0.8-billion-parameter model built to call tools on your own phone or laptop. It is small enough to run almost anywhere, and the interesting claim is that it beats models more than twice its size at the agent tasks it was trained for.
BotPsy starts from Qwen3.5-0.8B and is trained in four stages: supervised fine-tuning on tool-use conversations, distillation from three specialist teacher models of 4B parameters, and two rounds of reinforcement learning with rewards checked by rules. The aim is a personal assistant for simple tasks across search, travel, shopping, device settings, files, calendar and email. Tether says it will release open weights under Apache 2.0 in bf16, Q8 and Q4, as safetensors and GGUF files for its QVAC SDK and llama.cpp, which we review.
The reported gains are large for the size. On BFCL v4, a standard tool-calling benchmark, the score rises from 23.4 for the base model to 44.0; on ACEBench from 30.1 to 58.7; and on tau-bench from 23.2 to 51.5. It also beats Qwen3.5-2B-Think, 2.5 times its size, on five of six agent benchmarks, losing only on Claw-Eval, 48.9 to 46.5. Two details matter for everyday use. On tests of declining when no suitable tool exists, it scores 86.1, the best of those tested. And it reasons more efficiently: getting stuck in repetitive loops on IFBench falls from 22.9% to 1.4%.
For readers interested in private, offline assistants, this is a useful data point. A model this small can run through llama.cpp, Ollama or LM Studio on ordinary hardware, and keeping tool calls local means your calendar and messages need not leave the device. The trade-off is knowledge: BotPsy is designed to look facts up rather than know them, and Qwen3.5-2B-Think still scores far higher on general knowledge.
The limits are spelled out by Tether itself. All benchmarks are its own runs. The model is weak at multi-step research, scoring 15 on complex information retrieval tasks, and much weaker in Hindi and Russian than in English. Red-teaming still found unsafe responses 21% of the time. Tether says it is not intended for decisions involving money, health or personal information, and its licence wording limits it to research and educational use.