Tether, the entity behind USDT stablecoin, has introduced a medical AI model that not only fits in your pocket but also surpasses much larger competitors. Today’s launch of QVAC MedPsy by Tether’s AI Research Group marks a new era for medical language models designed to operate on smartphones and edge devices without requiring cloud services.
The standout feature is its 1.7 billion-parameter model, which outperforms Google’s MedGemma-4B on various medical benchmarks despite being less than half its size. On OpenAI’s HealthBench Hard—a test evaluating AI in realistic multi-turn clinical discussions graded by 262 physicians—MedPsy’s smaller model surpasses MedGemma-27B, a model nearly sixteen times larger.
Parameters represent the configurations and values a model learns during training. Theoretically, more parameters equate to better performance.
The evaluation spans from MedQA-USMLE, which assesses clinical knowledge through questions similar to US medical licensing exams scored for accuracy, to AfriMedQA, focusing on healthcare contexts in underserved African regions.
Tether’s CEO Paolo Ardoino attributes these gains to efficiency rather than sheer scale. “Our focus with QVAC MedPsy was enhancing model efficiency over increasing size,” he stated. “Our 4 billion parameter model outperforms those nearly seven times its size, while using up to three times fewer tokens per response.”
Token efficiency stands as another highlight. The 4B model averages about 909 tokens per response compared to 2,953 for similar systems—a reduction of 3.2x. Fewer tokens mean lower computational costs and faster responses, enabling local operation without cloud dependency.
“You can perform medical reasoning right where the data is—within a hospital’s system or on a device—without transferring sensitive information through the cloud or waiting on external processing,” Ardoino explained.
Available as quantized GGUF files (1.2 GB for 1.7 billion parameters and 2.6 GB for 4 billion), these models fit standard consumer hardware while maintaining performance, allowing healthcare facilities or individual practitioners to run them entirely on-device. This capability keeps patient records secure from third-party cloud infrastructures and potential HIPAA breaches.
While privacy features may appeal to some, using AI for medical opinions remains imperfect. An Oxford study published in February found LLMs often provide dangerous advice with incorrect answers and poor symptom handling. The researchers suggest AI can function as a “secretary, not physician.” Compounding the issue is compliance: Most current medical AIs route patient data through cloud servers, risking HIPAA exposure.
This release aligns with Tether’s recent activities. Last month, they launched the QVAC SDK, an open-source toolkit for building offline AI apps across various platforms. Previously, QVAC Health was introduced—a consumer wellness app that stores biometric data on-device. MedPsy is their first model specifically tailored for clinical reasoning.
The medical AI market currently stands at approximately $36 billion, with projections exceeding $500 billion by 2033, according to Tether’s announcement. Models and GGUF weights are now accessible at qvac.tether.io/models.