Tether unveils 800-million-parameter AI agent that beats larger rivals in company tests

Tether unveiled QVAC BotPsy on Wednesday as an AI agent for everyday tasks on personal devices. The model contains 800 million parameters. Tether's tests found that BotPsy surpassed competing models with up to 2.5 times as many parameters. BotPsy achieved the best results in five of six tool-use evaluations involving five models. Those results suggest that AI assistants may not require very large models to perform practical tasks.

BotPsy can handle multi-step operations involving phone settings, messaging and scheduling, potentially without transferring personal information to cloud servers. The agent determines which software tools a task requires, runs commands in sequence and adapts when an operation fails. Tether described a messaging test in which BotPsy encountered disabled Wi-Fi and a full inbox. Tether said BotPsy enabled Wi-Fi, freed sufficient storage and sent the message. The underlying model instead kept repeating an unsuccessful operation.

Tether built BotPsy using Alibaba's Qwen3.5-0.8B architecture and provided additional training focused on software-tool interaction. Parameters are numerical values learned during training, and models with fewer parameters generally require less memory. Higher parameter counts generally improve performance, and today's frontier AI models contain trillions of parameters.

Tether's technical documentation gives BotPsy a score of 44.0% on BFCL v4. BFCL v4 measures function selection and execution across tasks and conversations, using a weighted success percentage across tool-use categories. The underlying Qwen3.5-0.8B scored 23.4% on that benchmark. The larger Qwen3.5-2B achieved 37.7%.

BotPsy achieved 58.7% on ACEBench, which measures function-call accuracy, including tool selection and the arguments supplied to tools. The base model achieved 30.1% on ACEBench. The 2-billion-parameter competitor achieved 55.4%. Tether's six evaluations produced an average score of 45.0% for BotPsy. The larger Qwen model averaged 38.8% across those evaluations.

BotPsy extends QVAC, an AI ecosystem that Tether is developing around user-controlled devices rather than complete dependence on centralized cloud infrastructure. Tether, the company behind USDT, the world's largest stablecoin, already has local QVAC models for medical uses, translation and speech processing. Tether introduced QVAC in May 2025 as an alternative to AI services from major technology companies. Tether's longer-term goals included autonomous agents that could use applications and potentially make payments with Bitcoin and USDT.

Tether added a science and engineering training dataset containing 41 billion tokens in October 2025. Tether also introduced QVAC Workbench, an application for running AI models locally. Earlier this year, Tether released a software development kit for building offline AI applications across mobile and desktop operating systems. Tether also developed QVAC MedPsy for medical AI tasks on consumer hardware. The 1.7-billion-parameter MedPsy version surpassed Google's larger MedGemma-27B on HealthBench Hard, which assesses realistic clinical conversations against criteria developed by physicians.

BotPsy is available now as an open-weight model licensed under Apache 2.0, allowing developers to download, modify and incorporate the model into applications. Tether has also made compressed versions available to lower memory needs for local use. Tether has not provided a launch date for a finished consumer assistant using BotPsy. The published model card states that the release excludes training code and training data.