Tether, the company behind USDT, has introduced a new artificial intelligence agent named QVAC BotPsy. Despite operating with just 800 million parameters, BotPsy exceeded the performance of competing models with up to 2.5 times as many parameters on the majority of the tool-use benchmarks conducted in company trials.
Compact model surpasses expectations
BotPsy is designed to execute multi-step daily tasks on personal devices, including functions such as adjusting phone settings, sending messages, and managing schedules. According to Tether engineers, the model may be able to perform these actions locally without transferring user data to cloud servers, potentially enhancing privacy.
While leading frontier AI models now feature trillions of parameters, Tether demonstrated in its internal tests that BotPsy managed tasks more efficiently than larger language models in five out of six benchmark evaluations.
Distinguishing itself from traditional chatbots, BotPsy operates as an agentic system, meaning it can identify necessary tools, execute commands in sequence, and adapt to errors during operation. For instance, in one test case where an AI assistant had to send a message but faced disabled Wi-Fi and full storage, BotPsy enabled Wi-Fi, cleared space, and completed the task. Competing models failed to adjust and repeated unsuccessful attempts.
Tether reported that BotPsy scored 44.0% on the BFCL v4 benchmark, which assesses an AI model’s ability to select and execute software functions across conversations and tasks. The underlying Qwen3.5-0.8B architecture achieved 23.4%, and the larger Qwen3.5-2B reached 37.7%.
BotPsy also achieved 58.7% on ACEBench, which measures the precision of function calls. The base model recorded 30.1%, while the 2-billion-parameter competitor posted 55.4%.
| Model | Parameter Count | BFCL v4 | ACEBench | Avg. Tool-Use Score |
|---|---|---|---|---|
| BotPsy | 800M | 44.0% | 58.7% | 45.0% |
| Qwen3.5-0.8B | 800M | 23.4% | 30.1% | – |
| Qwen3.5-2B | 2B | 37.7% | 55.4% | 38.8% |
Mini dictionary: Qwen3.5-0.8B, developed by Alibaba, is a compact large language model architecture specifically designed for efficiency and local device deployment, enabling complex AI tasks with fewer resources than larger models.
BotPsy recorded the highest scores in five out of six tool-use benchmarks, maintaining an average performance of 45.0% in Tether’s internal tests, compared to 38.8% for larger models such as Qwen3.5-2B.
Growth of Tether’s QVAC AI ecosystem
The launch of BotPsy expands Tether’s QVAC initiative, an AI ecosystem emphasizing localized computation. QVAC is intended as an open alternative to cloud-based AI services offered by major technology firms.
Tether first announced QVAC in May 2025, outlining plans for user-controlled agents capable of directly interacting with applications and, in the future, possibly handling autonomous payments using Bitcoin and USDT.
The company expanded the project in October 2025, introducing a 41-billion-token dataset for science and engineering AI projects and unveiling QVAC Workbench for running AI models locally. Additionally, a software development kit (SDK) was released to help developers build offline AI applications for both mobile and desktop platforms.
In the medical sector, Tether introduced QVAC MedPsy, a local AI system for healthcare applications. Its 1.7-billion-parameter version outperformed Google’s MedGemma-27B on the HealthBench Hard benchmark, which evaluates AI systems with physician-based criteria in clinical scenarios.
Mini dictionary: QVAC is Tether’s expanding AI suite focused on decentralized, privacy-oriented AI solutions that operate directly on users’ devices without mandatory cloud reliance.
Open access and future development
BotPsy is available as an open-weight model under the Apache 2.0 license, allowing developers to download, modify, and incorporate it into their applications. Tether also released compressed versions to enable deployment on devices with limited memory.
A release date for a complete consumer-facing assistant based on BotPsy has not been specified. Tether’s documentation clarified that neither the training code nor the dataset used for BotPsy’s development are included with the current release.
BotPsy can be integrated or adapted by developers due to its open-weight access, but user-ready assistants and the full training details remain unavailable.
