Tether’s 1.7B MedPsy runs on phones, tops Google models

Tether’s AI Research Group released QVAC MedPsy on Thursday, a 1.7 billion-parameter medical language model designed to run on smartphones and other edge devices. Tether reported that the 1.7B model outscored Google’s MedGemma-4B and beat MedGemma-27B on OpenAI’s HealthBench Hard, a benchmark of multi-turn clinical conversations graded by 262 physicians.

The company also published a 4 billion-parameter variant. Both models are intended to operate without cloud infrastructure and to run locally on consumer hardware or hospital servers. HealthBench Hard includes multiple tests, such as MedQA-USMLE, which measures clinical knowledge with licensing-exam-style questions, and AfriMedQA, which evaluates performance in African healthcare contexts.

Tether reported that the 4B model generates about 909 tokens per response, compared with roughly 2,953 tokens for comparable systems, a 3.2-times reduction. The company said lower token counts reduce compute demands and shorten response times, which supports deployment in clinics, hospitals and remote settings with limited cloud access.

The models are distributed as quantized GGUF files that compress core weights while preserving most benchmark performance. The 1.7B variant is supplied in a 1.2 GB file and the 4B variant in a 2.6 GB file. Tether said those file sizes allow the models to run entirely on-device, reducing the need to route patient data through third-party cloud servers and potentially lowering HIPAA exposure associated with cloud-based queries. The weights and model files are available at qvac.tether.io/models.

Tether framed the release alongside other recent products, including a QVAC SDK for building local, offline apps across mobile and desktop platforms and QVAC Health, a wellness app that keeps biometric data on-device. The company described the current medical AI market as roughly $36 billion and included a projection to $500 billion by 2033 in its announcement.

Independent research has found that large language models can produce incorrect or misleading medical guidance, including dangerous recommendations and poor handling of nuanced symptoms. Paolo Ardoino, Tether’s chief executive, commented on efficiency in a statement: “With QVAC MedPsy, our focus was improving efficiency at the model level, rather than scaling up size. Our 4 billion model exceeded results from models nearly seven times its size, while using up to three times fewer tokens per response.” Tether said the on-device design is intended to limit data exposure while acknowledging the broader limits of medical AI.