A language model just 280 megabytes in size, running directly on an Android smartphone, can carry on a therapeutic conversation at 17 tokens per second — without a single byte of information ever leaving the device. This is not a concept deck or a conference slide: the system, called MoPHES, was built by researchers Wei, Zhou, and Wang and published in an IEEE journal in 2025.

Why Is Privacy the Central Problem in Digital Therapy?

According to the WHO, more than 70% of people with mental health conditions never seek help. Among the reasons are stigma and the fear that personal information will leak. That fear is well-founded: even in research settings, ethics boards prohibit sharing real therapeutic session data for analysis. If privacy cannot be guaranteed under laboratory conditions, what can we expect from consumer apps?

Traditional online mental health services, including chatbots that have demonstrated effectiveness in clinical trials, rely on the cloud. Every message travels to a remote server, gets processed, and comes back. Even with end-to-end encryption, the data is stored somewhere — and could theoretically be compromised.

What Is an On-Device Model and How Does It Work?

On-device means exactly what it sounds like: the model runs on your phone. No server, no cloud, no internet connection. MoPHES uses two compact language modules of 0.5 billion parameters each, executed through the llama.cpp framework. After Q4_K_M quantization, each model takes up about 280 MB — less than an average mobile game.

On the test device, a Xiaomi 13 Ultra (8 cores, 16 GB of RAM), the system generates conversational responses at 17.3 tokens per second. A mental state assessment takes 4.2 seconds. That is a comfortable speed — the user does not feel any lag.

Why Two Modules Instead of One?

The MoPHES architecture separates concerns. One module handles dialogue — it responds to the user's messages, asks clarifying questions, and applies supportive communication techniques. The second module acts as an analyst: it evaluates the user's mental state throughout the conversation and saves the results to a local configuration file on the device.

This separation matters: the conversational model can be empathetic and flexible in its wording, while the analytical model stays rigorous and structured. The agent retrieves session history from local memory to personalize each subsequent conversation. All of this happens without a single server call.

What Does This Mean for Nearly a Billion People?

According to WHO estimates, nearly a billion people worldwide need mental health support. Most of them do not receive it — due to a shortage of professionals, the cost of therapy, geographic remoteness, or fear of judgment. Mental health chatbots have already proven effective for at least mild to moderate symptoms (Abd-Alrazaq et al., 2020).

But trust remains a bottleneck. A study by Song and colleagues (2024) found that users are willing to open up to an AI conversational partner, but only when they are confident their words will not be read by a third party. On-device models remove this barrier technically, not just legally — the data simply never leaves the device.

What Are the Limitations of the On-Device Approach?

It would be dishonest to gloss over the boundaries. Models with 0.5 billion parameters are significantly less capable than their cloud-based counterparts in terms of depth and flexibility of responses. They handle structured tasks well — screening, protocol-based supportive dialogue — but they are not yet sufficient for complex psychotherapeutic work.

Moreover, not every smartphone has 16 GB of RAM. For mass adoption, even more compact models or a hybrid approach are needed: basic functions on the device, with advanced capabilities in the cloud with the user's consent. It is also worth remembering that digital monitoring tools raise their own questions about the boundaries of data collection.

What Comes Next?

MoPHES is the first fully autonomous AI mental health support system that runs on a mobile device. It demonstrates that privacy and accessibility do not have to be in conflict. As quantization techniques and mobile chip optimization continue to advance, on-device models will become even more compact and accurate.

Already today, services like Nearby use evidence-based approaches to mental health support. And as ethical standards for AI in psychotherapy become clearer, the line between a laboratory experiment and an everyday self-care tool continues to blur.

Frequently Asked Questions

Can an AI on a phone replace a psychotherapist?

No. On-device models are suitable for supportive conversation, screening, and mood monitoring, but not for full psychotherapy. They complement, rather than replace, work with a human professional.

How accurate are compact models compared to GPT-4 and similar systems?

Models with 0.5 billion parameters fall noticeably short of large cloud-based models in open-ended text generation. However, for narrowly defined tasks — structured mood assessment, protocol-based supportive responses — their accuracy is sufficient for practical use.

Is it true that no data is sent anywhere at all?

In the MoPHES architecture, yes. The model runs fully offline, and all records are stored locally. However, each specific app may implement this architecture differently, so it is always worth checking the service's privacy policy.

What kind of smartphone is needed to run such a model?

The researchers tested on a Xiaomi 13 Ultra with 16 GB of RAM. For comfortable performance, a device with 8+ GB of RAM and a modern processor is recommended. As models are optimized further, the requirements will decrease.