On-device psychiatric AI: Gemma, Phi, and Qwen run without sending data to the cloud

A research team led by Eranga Bandara with 11 collaborators published a paper on April 20, 2026 on arXiv describing a mobile application for psychiatric support that runs entirely locally, without sending any data to the cloud.

Three small models instead of one large one

The system uses three quantized and fine-tuned language models: Gemma, Phi-3.5-mini, and Qwen2. All run locally on the mobile device, and an orchestration layer coordinates their ensemble inference and consensus reasoning. Instead of relying on a single large cloud model, the smaller models complement and cross-check each other, increasing the reliability of diagnostic assessments. arXiv is an open archive of scientific preprints operated by Cornell University. Initial evaluation shows that the on-device version achieves accuracy comparable to cloud-based versions with real-time latency on standard mobile hardware.

DSM-5 assessments and clinical application

The application generates assessments aligned with the DSM-5 classification of the American Psychiatric Association. There are two main functions: supporting clinicians in differential diagnosis and symptom mapping, and patient self-screening with safety mechanisms that direct the user toward professional help. The authors emphasize that no patient data leaves the device, which is critical for the consent of sensitive populations.

Why privacy is critical

The paper focuses on contexts where fear of data exposure deters people from seeking help. Examples include the military, where a psychiatric diagnosis can affect security clearances; the criminal justice system, where inmates do not trust centralized databases; and remote or rural healthcare environments where there is no stable internet connection. The zero-egress approach technically eliminates this risk because data is never available to third parties and cannot be intercepted on the way to an external server. An additional advantage is resilience to network outages, which is important in military and rural scenarios. The authors note that model quantization is key to practical usability because it enables execution on mid-range phones without overheating and excessive battery drain. The paper is published under the Creative Commons Attribution 4.0 license, encouraging open scientific validation of the method.