Tether launches QVAC MedPsy with 1.7 billion parameters, surpassing MedGemma-27B on HealthBench Hard, using three times less computing resources and requiring no cloud infrastructure.
Tether , the company famous for USDT stablecoin, has just announced QVAC MedPsy, a line of medical AI models designed to run directly on smartphones, wearables, and edge devices without cloud connectivity. The most noteworthy aspect isn't the product's unusual origin—a stablecoin company venturing into medical AI—but rather the accompanying performance figures.
On HealthBench Hard, OpenAI's benchmark for evaluating AI through multiple rounds of clinical conversations scored by 262 doctors, Tether 's 1.7 billion-parameter model outperformed Google's MedGemma-27B, a model nearly 16 times larger in scale.
The 4 billion parameter model also outperformed its larger competitors by nearly seven times while using an Medium of only 909 Token per response, compared to 2,953 Token in comparable systems, equivalent to a 3.2-fold reduction in computational cost.
The advantages of Token efficiency are not just technical numbers. Fewer Token mean faster response times, lower costs, and most importantly, the ability to operate entirely locally on mainstream consumer hardware.
The models were released as quantified GGUF files with sizes of 1.2GB and 2.6GB, respectively, small enough to install on mobile devices while maintaining much of the performance across assessment suites from MedQA-USMLE to AfriMedQA, a set of standards focused on the resource-scarce healthcare landscape in Africa.
Data privacy — a real competitive advantage
Tether CEO Paolo Ardoino positions QVAC MedPsy not as a competition of scale but as a matter of efficiency: medical inference right where the data exists, within the hospital system or on a personal device, without transferring sensitive information to a third-party infrastructure.
Given that most current medical AI processes patient data via cloud servers and creates risks related to HIPAA regulations, this argument carries practical weight for remote hospital and clinic systems.
However, technical potential does not equate to clinical readiness. A February study by Oxford University showed that large language models still frequently give inaccurate medical advice and poorly handle complex symptoms, leading researchers to conclude that AI should play the Vai of "secretary, not doctor."
QVAC MedPsy was released amidst a healthcare AI market currently valued at approximately $36 billion and projected to exceed $500 billion by 2033—a space large enough for device-based models to find their place, even if they don't yet replace the role of clinicians.
