Tether's new medical AI runs on 3x less compute than rivals

Tether, the company behind the USDT stablecoin, has released a new set of medical artificial intelligence models that challenge the industry’s focus on scale, with its most efficient system beating Google models nearly seven times its size on key benchmarks. The QVAC MedPsy models are designed to run locally on devices like smartphones, a move that could sidestep major privacy risks in the growing healthcare AI sector.

"With QVAC MedPsy, our focus was improving efficiency at the model level, rather than scaling up size," Tether CEO Paolo Ardoino said in a statement. "You can run medical reasoning where the data already exists, inside a hospital system or on a device, without moving sensitive information through the cloud or waiting on external processing."

The company released two models, with the 4 billion-parameter version scoring 70.54 across a suite of eight medical benchmarks, surpassing Google’s MedGemma-27B, a model almost seven times larger. A smaller 1.7 billion-parameter model also outperformed a comparable Google model by over 11 points. Tether credits the performance to a specialized training process that combines supervision, curated clinical reasoning data, and reinforcement learning. The 4B model is also three times more efficient, generating responses in roughly 909 tokens compared to about 2,953 for similar systems.

The release targets a critical vulnerability for AI in healthcare: patient privacy. The healthcare AI market is projected to expand from around $36 billion today to over $500 billion by 2033, but most systems rely on the cloud. This requires transmitting sensitive patient records to external servers, creating compliance and privacy risks under regulations like HIPAA. By running entirely on-device, QVAC MedPsy keeps all data local.

On-Device AI as a Privacy Solution

The architecture of Tether's models offers a direct solution to the privacy concerns that have accompanied the rollout of AI in clinical settings. While doctors have started using AI to summarize patient visits and reduce documentation time, the process often involves uploading conversations to the cloud, as some physicians at Orlando Health have done. This has led to patient apprehension over data security, even with assurances of HIPAA compliance.

Tether’s approach eliminates that step. The models are being released in a compressed GGUF format, with the smaller version taking up just 1.2 GB, allowing it to be installed directly on standard hospital hardware or a doctor's smartphone. This means a rural clinic or individual practitioner could use the AI without needing high-speed internet or a cloud subscription, and without patient data ever leaving the premises.

A Secretary, Not a Physician

Despite the performance gains and privacy advantages, the role of large language models in medicine remains a subject of intense debate. An Oxford study published in February found that AI models frequently provide wrong answers, confused guidance, and dangerous medical advice when handling nuanced symptoms. The researchers concluded that current AI has a role as a "secretary, not physician."

This release is part of a broader strategic push by Tether into artificial intelligence. The company recently launched the QVAC SDK, an open-source toolkit for building local, offline AI applications, and QVAC Health, a consumer wellness app that keeps biometric data on-device. The QVAC MedPsy models, now available on the open-source AI platform Hugging Face, are the first from the company specifically trained for clinical reasoning.

This article is for informational purposes only and does not constitute investment advice.