Tether'S Medical AI Beats Models 16x Bigger

@tether's AI Research Group launched @qvac MedPsy on Wednesday, releasing a family of medical language models compact enough to run on a smartphone yet capable of outperforming rivals many times their size on clinical tests.

Benchmark Results That Challenge the Scale-First Assumption

MedPsy comes in two sizes, 1.7B and 4B parameters, and is purpose-built for edge deployment through the QVAC ecosystem. The smaller 1.7B model beat Google's MedGemma-4B by 11.42 points across seven medical tests. On the more demanding HealthBench Hard assessment, it surpassed MedGemma-27B, a model 16 times its size. The 4B version scored 70.54 on closed-ended medical benchmarks, surpassing MedGemma-27B despite being nearly seven times smaller, with the gap widening further on realistic health scenarios including HealthBench Hard (58.00 vs 42.00) and MedXpertQA (30.61 vs 25.18). The 4B model also generated responses using 3.2 times fewer tokens, pointing to meaningful gains in computational efficiency alongside raw accuracy.

The current landscape of medical large language models presents a stark trade-off between capability and deployability. Google's MedGemma-27B delivers strong benchmark performance, but at 27 billion parameters it is entirely infeasible for edge deployment, requiring GPUs with tens of gigabytes of VRAM. Tether's approach bets on curated training methodology over brute-force scale.

Why On-Device Deployment Matters for Healthcare

The models ship in quantized formats of 1.2 GB and 2.6 GB, small enough to run on a smartphone or a hospital workstation without an internet connection. Medical data is uniquely sensitive. Patient records, diagnostic queries, and clinical notes contain protected health information governed by strict regulatory frameworks including HIPAA in the United States and GDPR in Europe. The dominant paradigm of cloud-hosted medical AI requires this data to leave the user's device and be processed on remote servers, creating compliance burdens and a fundamental tension between AI capability and patient privacy. Running inference locally removes that barrier entirely.

CEO @paoloardoino framed the release as a deliberate choice of efficiency over scale, consistent with Tether's broader push into privacy-preserving, on-device AI through the QVAC platform. The MedPsy models are available under the Apache 2.0 license for research and educational purposes, and are specifically designed for deployment in bandwidth-constrained environments, privacy-sensitive clinical workflows, and low-resource healthcare settings where data must never leave the device.

The launch arrives at a significant moment for the sector. The global AI in healthcare market was valued at $36.67 billion in 2025 and is projected to reach $505.59 billion by 2033, growing at a compound annual growth rate of 38.90%. Compliance with data-privacy regulations remains one of the most cited obstacles to broader adoption, which is precisely the problem MedPsy is designed to solve.

Sources:
QVAC MedPsy: State-of-the-Art Medical and Healthcare Language Models for Edge Devices, Hugging Face
Tether Data Introduces QVAC Fabric LLM, Tether.io
AI in Healthcare Market Size and Share Report, Grand View Research