Tether’s Push Into Edge AI Yields Powerful Mini

Agent That Beats Larger Rivals

Tether, the financial powerhouse renowned globally as the issuer of USDT—the world’s largest and most widely traded stablecoin—is increasingly proving its capability to engineer remarkably efficient, lightweight artificial intelligence models. Moving well beyond its traditional crypto-native footprint, the company has unveiled an innovative AI agent designed to operate directly on everyday personal devices like smartphones, challenging the industry’s longstanding reliance on massive cloud-based architectures.

The newly introduced system, known as QVAC BotPsy, is an 800-million-parameter AI agent built specifically to handle functional, everyday tasks directly on consumer hardware. In the field of artificial intelligence, parameters function similarly to adjustable neural pathways or numerical knobs that dictate a model’s capacity for processing information and executing commands; as a general rule across the industry, a higher parameter count correlates directly with enhanced performance and reasoning capabilities. While contemporary frontier AI systems developed by major technology conglomerates often operate on scales reaching into the trillions of parameters, Tether’s latest innovation achieves outsized results at a fraction of the computational footprint.

According to extensive testing conducted by Tether, this relatively diminutive model successfully outperformed competing systems possessing up to 2.5 times as many parameters. This stark performance differential suggests a compelling shift in the trajectory of consumer technology: AI assistants may not actually require colossal, resource-heavy neural networks to effectively accomplish practical, real-world tasks.

Unlike traditional conversational chatbots that simply process prompts and generate text-based replies in a static window, BotPsy is fundamentally designed as an agentic system. This means it is engineered to take autonomous, multi-step actions within software environments. Specifically, the agent can intelligently identify which tools a given application requires, execute complex commands in the correct sequence, and dynamically adjust its strategy when unexpected errors or roadblocks arise during execution.

To illustrate this practical utility, Tether outlined a realistic benchmark scenario involving an AI assistant tasked with sending an urgent message on a mobile device, only to encounter two simultaneous technical hurdles: the device’s Wi-Fi connection was disabled, and the internal storage inbox was entirely full. While a standard baseline model repeatedly attempted the exact same failed operation without recognizing the environment, BotPsy autonomously intervened by enabling the Wi-Fi connection, clearing out sufficient internal storage space, and successfully completing the primary messaging task.

While such operational achievements may appear mundane on the surface, they represent precisely the kind of contextual problem-solving capabilities that render AI agents genuinely useful beyond a confined browser window or chatbot interface.

The structural foundation of BotPsy is built upon Alibaba’s Qwen3.5-0.8B architecture, an already compact language model that Tether further trained and fine-tuned to specialize in software tool interactions and digital environment navigation. Because an 800-million-parameter model requires significantly less active memory and processing power to load and execute, it is uniquely suited for local deployment on consumer-grade hardware without draining device resources.

Technical documentation published by Tether highlights BotPsy’s impressive performance across standardized industry evaluations. On BFCL v4—a rigorous benchmark that evaluates an AI system’s aptitude for selecting and executing software functions across diverse tasks and conversational contexts—BotPsy achieved a score of 44.0%. Scores on this evaluation reflect a weighted percentage of successful performance across specialized tool-use categories. By comparison, the underlying Qwen3.5-0.8B base model managed a score of 23.4%, while the considerably larger Qwen3.5-2B model reached 37.7%, demonstrating that Tether’s specialized training methodology yielded substantial gains over the unoptimized architecture.

Furthermore, BotPsy secured a 58.7% score on ACEBench, a benchmark designed to measure the precision of function calls, including the correct selection of appropriate tools and corresponding arguments. On this same test, the base model achieved 30.1%, while the larger 2-billion-parameter competitor lagged slightly behind BotPsy with a score of 55.4%. Across a comprehensive evaluation testing five distinct models over six separate tool-use assessments, BotPsy posted the highest results in five of those categories, registering an average score of 45.0% compared to 38.8% for the larger Qwen model.

BotPsy represents the latest milestone within QVAC, Tether’s broader strategic initiative aimed at constructing a decentralized AI ecosystem capable of running locally on user-controlled hardware rather than depending entirely on centralized cloud infrastructure operated by big tech firms. Tether originally introduced the QVAC project in May 2025, positioning the initiative as a viable counterweight to corporate AI services. From its inception, the project’s long-term roadmap envisioned autonomous agents capable of interacting seamlessly with a wide array of software applications and potentially facilitating native peer-to-peer payments utilizing Bitcoin and USDT.

The company steadily expanded this foundational initiative in October 2025 by releasing a massive 41-billion-token dataset explicitly curated for training science and engineering models. This release was accompanied by QVAC Workbench, a dedicated application allowing users to run complex AI models locally on their own machines. Earlier this year, Tether continued its developer outreach by launching a specialized software development kit, providing engineers and builders with the necessary tools to construct offline-first AI applications capable of functioning smoothly across various mobile and desktop operating systems.

In addition to BotPsy, Tether has developed specialized models tailored for specific industries, such as QVAC MedPsy, a medical AI system likewise engineered to run efficiently on standard consumer hardware. Notably, a 1.7-billion-parameter iteration of MedPsy outperformed Google’s much larger MedGemma-27B model on HealthBench Hard, a demanding evaluation of realistic clinical conversations judged against rigorous, physician-developed criteria.

Currently, BotPsy is available to the broader developer community as an open-weight model distributed under the permissive Apache 2.0 license. This licensing structure allows developers to freely download, modify, and integrate the technology into their own proprietary applications. Additionally, Tether has released compressed versions of the model specifically engineered to minimize memory overhead for local deployment on resource-constrained devices.

Tether has not yet announced a definitive commercial release date for a finished, consumer-facing assistant powered directly by BotPsy. Furthermore, the company’s published model card notes that the underlying training code and proprietary training datasets are not included in the current open-weight release.