qvac/MedPsy-4B

MedPsy-4B: A Compact, High-Performance Medical LLM

MedPsy-4B, developed by Tether AI Research, is a 4 billion parameter, text-only medical and healthcare language model. Built upon the Qwen3-4B-Thinking-2507 base, it undergoes a multi-stage post-training process involving supervised fine-tuning (SFT) and reinforcement learning (RL) on specialized medical datasets.

Key Capabilities & Differentiators

• Exceptional Medical Performance: Achieves a score of 70.54 on closed-ended medical benchmarks, outperforming models nearly 7x its size, such as MedGemma-27B-text-it (69.95).
• Real-World Clinical Strength: Scores 74.00 on HealthBench and 58.00 on HealthBench Hard, significantly surpassing MedGemma-27B.
• High Token Efficiency: Demonstrates a 3.2x reduction in average response length compared to its backbone model, leading to faster inference, lower compute costs, and reduced latency.
• Privacy-First Design: Supports fully on-device inference via the QVAC SDK, ensuring patient data remains on the device.

Intended Use Cases

• Research: Ideal for medical language understanding and reasoning studies.
• Developer Tools: Suitable for building prototypes and tools for health-related applications.
• On-Device Retrieval: Excellent for privacy-sensitive medical information retrieval in edge environments.

Limitations

• Not a Medical Substitute: Outputs are not a substitute for professional medical judgment or clinical diagnosis.
• Hallucinations: May generate plausible but incorrect medical information.
• English Only: Performance is validated primarily in English.
• Text Only: Cannot interpret non-text medical data like images or lab results.
• Knowledge Cutoff: Does not reflect the latest medical guidelines or evidence.