Tether's Tiny AI Agent Beats Models 2.5 Times Its Size at Controlling Your Phone

Tether, the company behind the world's largest stablecoin, USDT, is getting surprisingly good at making tiny AI models punch above their weight.

The company on Wednesday unveiled QVAC BotPsy, an 800-million-parameter AI agent designed to carry out everyday tasks on personal devices. Parameters are essentially the knobs that determine how much an AI model can do, and more of them generally translate to improved performance. And for reference, frontier AI models today run in the trillions of parameters.

In the company's own testing, Tether’s relatively tiny model outperformed competitors with up to 2.5 times as many parameters, suggesting that AI assistants may not need enormous models to get things done.



Unlike traditional chatbots that simply respond to questions, BotPsy is agentic, meaning it’s designed to take action. It can identify which tools an application needs, execute commands in the right order, and adjust when something goes wrong.

For example, Tether described a benchmark scenario in which an AI assistant needed to send a message but encountered two problems: Wi-Fi was disabled and the inbox was full. BotPsy turned on Wi-Fi, cleared enough storage, and completed the task, while the underlying model repeatedly attempted the same failed operation.

That's a fairly mundane achievement, but it's precisely the kind of problem that makes AI agents useful outside a chatbot window.

BotPsy is built on Alibaba's Qwen3.5-0.8B architecture, a relatively small language model that Tether trained further to specialize in interacting with software tools. Its 800 million parameters represent the numerical values the model learned during training, and a smaller parameter count generally means less memory is needed to run it.

According to Tether's technical documentation, BotPsy scored 44.0% on BFCL v4, a benchmark that evaluates an AI system's ability to select and execute software functions across different tasks and conversations. Scores reflect a weighted percentage of successful performance across tool-use categories.

The underlying Qwen3.5-0.8B model scored 23.4%, while the considerably larger Qwen3.5-2B reached 37.7%.

BotPsy also scored 58.7% on ACEBench, which evaluates the accuracy of function calls, including selecting the appropriate tools and arguments. Its base model managed 30.1%, while the 2-billion-parameter competitor scored 55.4%.