Tether Releases QVAC BotPsy, an 800M-Parameter On-Device AI

Tether said on Nov. 11 that it has released QVAC BotPsy, an open-weight AI model with 800 million parameters (BotPsy-0.8B). It is built to run locally on a phone or laptop and handle agentic tasks: calling tools, planning steps and declining requests it cannot fulfil. Tether says it beat models up to 2.5 times its size across six benchmarks.

Details

The weights are published on Hugging Face, so developers can download and fine-tune them. All figures below are Tether's own test results; no independent verification has been published yet.

• On BFCL multi-turn (multi-step tool calls), BotPsy scored 46.1 versus 23.8 for Alibaba's Qwen3.5⁠–⁠2B.
• On BFCL v4 it scored 44.0 versus 23.4 for its base architecture, on ACEBench 58.7 versus 30.1, and on τ-bench 51.5 versus 23.2.
• On the BFCL hallucination test, where the model must recognize that no suitable tool exists and refuse, it scored 86.1.

The work comes from Tether AI Research under the QVAC project. CEO Paolo Ardoino framed the local approach as a privacy benefit: an assistant that manages your calendar, messages and documents needs access to personal data, and running AI locally can reduce the need to send it to centralized providers.

What it means for the market

This is not a trading event: the release has no direct bearing on USDT or other stablecoins. It does show the issuer of the largest stablecoin (see the Tether page) pushing further into AI, an area where crypto companies keep making headlines. How the model performs in practice will become clearer once developers test it independently.

For current market data, see the daily brief.