Tether Unveils BotPsy, a Tiny AI Agent Outperforming Larger Models

The company behind USDT, known as Tether's stablecoin, has developed a compact artificial intelligence system that demonstrates superior performance compared to much larger rivals. On Wednesday, the organization introduced QVAC BotPsy, an AI agent featuring 800 million parameters built to manage daily tasks directly on personal hardware.

Parameters act as the computational variables governing model capability, with frontier technology typically operating in the trillions. However, internal testing revealed that this relatively miniature architecture outperformed competitors possessing up to 2.5 times more parameters, indicating that efficient performance does not strictly require massive scale.

Understanding the BotPsy Architecture

Unlike conventional conversational bots, BotPsy operates as an agentic system engineered for execution. It determines tool requirements, runs commands sequentially, and adapts when errors occur. In a demonstration scenario involving a disabled internet connection and a full inbox, the agent successfully restored connectivity and cleared storage space to finish the objective, whereas the underlying foundation model repeatedly failed.

The system utilizes Alibaba's Qwen3.5-0.8B foundation, which underwent specialized training by the creators to enhance software interaction capabilities. Technical documentation shows the agent achieved 44.0 percent on the BFCL v4 benchmark evaluating function execution, surpassing both its base version at 23.4 percent and a larger 2-billion-parameter alternative at 37.7 percent. Furthermore, it reached 58.7 percent on ACEBench for function call accuracy.

BotPsy packs 800 million parameters into an AI assistant that outperforms larger rivals locally.

This release forms part of QVAC, an initiative launched in May 2025 to create a decentralized artificial intelligence ecosystem running on user-controlled hardware rather than centralized cloud infrastructure. The project previously introduced science datasets, local execution workbenches, software developer kits, and a medical assistant named MedPsy.

Developers can access BotPsy as an open-weight model distributed under the Apache 2.0 license, complete with compressed variations designed to minimize memory consumption during offline deployment. Training data and code remain excluded from the distribution, and a consumer-ready application launch date has not been shared.

Why it matters

This development signals a shift toward efficient, locally hosted artificial intelligence running directly on consumer hardware rather than relying on centralized cloud infrastructure. By proving that smaller models can surpass larger competitors in tool execution, the project expands decentralized software capabilities for developers worldwide.