Tether Releases BotPsy, a 0.8-Billion-Parameter Agent Model Built to Run on Devices

BotPsy is built on Alibaba's Qwen3.5-0.8B and is pitched as a personal assistant for tasks of low to medium difficulty: web search, travel bookings, commerce and finance, customer operations, device and vehicle control, files, calendars, email and messaging. The team, which publishes as Tether AI Research under Tether Data, S.A. de C.V., said in a Hugging Face blog post that the model will be released with open weights under the Apache 2.0 licence for research and educational use, in bf16, Q8 and Q4 precisions and as GGUF files for its QVAC SDK and llama.cpp.

The release is a bet against routing every task to frontier cloud models. The researchers argued that cloud models carry high and unpredictable operating costs, that providers can change, deny or withdraw access at any time, and that sending data to them raises privacy and control risks, while many routine workloads can be handled by small specialised models running locally.

According to the team's own figures, BotPsy roughly doubles its base model's agentic scores: BFCL v4 rises from 23.4 to 44.0, ACEBench from 30.1 to 58.7 and τ-bench from 23.2 to 51.5. It also beat Qwen3.5-2B-Think, a model 2.5 times its size, on five of six tool-use benchmarks, losing only on Claw-Eval, 48.9 to 46.5. On τ²-bench's telecom customer-service test it scored 82.5 against 69.3 for the larger Qwen model.

The researchers put weight on knowing when not to act. On BFCL v4's hallucination tests, which check whether a model declines when no suitable tool exists, BotPsy scored 86.1, the highest of the models tested. Instruction following improved sharply, with IFEval rising from 59.7 to 90.8.

In automated red-teaming across six harm categories, BotPsy produced unsafe responses 21 per cent of the time, compared with 36 per cent for its base model, and 31 per cent under prompt-injection attacks, down from 46 per cent. The team also said the model reaches correct answers with about three times fewer reasoning tokens than its base on GPQA Diamond, and that runs stuck in repetitive reasoning loops on IFBench fell from 22.9 per cent to 1.4 per cent, which should cut latency on phones and laptops.

All the benchmark results are self-reported and have not been independently verified. The team cautioned that BotPsy still makes mistakes and is not intended for high-stakes use, and said future work will focus on expanding on-device agentic capabilities.