A model with just 500 million parameters outscored GPT-4.1 on the ROUGE-1 metric in therapeutic dialogues — 41.32 versus 40.04. This comes from the MoPHES study published in IEEE in October 2025. The authors — Wei, Zhou, and Wang — demonstrated that in psychological support, what matters isn't model size but the quality of training data.

What Is MoPHES?

MoPHES (Mobile Psychological Health Evaluation and Support) is a system built on MiniCPM4-0.5B, a language model fine-tuned specifically for conducting multi-turn therapeutic conversations. The key word here is "specifically." Instead of training a massive model on everything under the sun, the researchers took a compact model and fine-tuned it on a carefully curated corpus of psychological consultations.

The corpus was assembled from two Chinese datasets — PsyQA and EmoLLM. The original 113,552 question-answer pairs were filtered and transformed into 34,827 multi-turn dialogues simulating real consultations. Topics covered: family and marriage (50.6%), emotional issues (24.7%), and personal growth (13.4%).

Why Does a Small Model Beat a Large One?

General-purpose models like ChatGPT and GPT-4.1 are trained on trillions of tokens from the internet. They know a bit of everything — and nothing deeply. In a psychological context, this shows up in specific ways: they give advice instead of listening, repeat the same phrases, and struggle to maintain emotional context across long conversations.

The fine-tuned MiniCPM4-0.5B learned to do something different — to behave like a counselor, not an encyclopedia. On the ROUGE-1 metric, it scored 41.32 in the label strategy, while GPT-4.1 scored 40.04. This means the smaller model's responses more closely matched reference therapeutic replies in both content and vocabulary.

In manual expert evaluation — measuring understanding, empathy, professionalism, helpfulness, and safety — MoPHES scored 7.204 out of 10 in the label strategy. GPT-4.1 scored 8.685. The gap exists, but MoPHES became the top performer among all non-commercial models. Considering that GPT-4.1 is a product backed by billions of dollars, the results from a 0.5B-parameter model are impressive.

Why Did "Reasoning" Models Fail?

The study's most surprising finding: DeepSeek-R1-7B — a reasoning model optimized for logical deduction — produced the worst results among all tested systems. This is counterintuitive: you'd expect a model built for reasoning to better analyze a client's problems.

But therapy isn't a logic puzzle. A person sharing their pain doesn't need a step-by-step breakdown of the situation. They need to feel heard. Models designed for chain-of-thought reasoning literally "think out loud" instead of offering support. They're optimized to find the right answer — and in therapy, there often is no right answer.

What Does This Mean for the Future of AI Therapy?

Several takeaways worth remembering.

Accessibility. MoPHES was trained on a single A100 GPU. That's not a supercomputer — it's standard hardware available in the cloud for tens of dollars per hour. If a high-quality therapeutic model can be built without Google-scale infrastructure, the barrier to entry for mental health app developers drops dramatically.

Privacy. A 500-million-parameter model can run directly on a smartphone — without sending data to a server. For mental health support, this is critical: people are more likely to seek help when they're confident their words aren't being sent to the cloud.

Specialization beats scale. Research in recent years — SMILE, MeChat (2023), SoulChat (2023) — has already shown that synthetic and curated datasets for training therapeutic models produce strong results. MoPHES confirmed the trend: narrow specialization wins over generality.

Where Is the Line?

It's important not to confuse progress with readiness. MoPHES was trained on Chinese-language data — transferring it to other languages and cultural contexts will require separate work. Manual evaluation still gives the edge to commercial models for empathy and professionalism. None of the tested systems have undergone clinical trials — unlike Therabot, which reduced depression symptoms by 51%.

According to the WHO (2022), one in eight people worldwide lives with a mental health condition, and 75% of people in low-income countries receive no treatment at all. Compact specialized models are one realistic path toward closing that gap.

The Nearby project is built on exactly this logic: not chasing model size, but building a support system that understands context, maintains empathy, and operates within evidence-based frameworks.

Frequently Asked Questions

Can a 500M-parameter AI model replace a human therapist? No. MoPHES and similar systems are support tools, not replacements for professionals. They can help between sessions, in areas without access to therapists, or as a first step for people who aren't yet ready to talk to a human.

Why does it matter that the model is small? Compact models can run locally — on a phone or laptop — without an internet connection. This protects privacy and makes support accessible even in regions with poor network coverage.

How does a fine-tuned model differ from ChatGPT playing "therapist"? ChatGPT and GPT-4.1 are general-purpose models that adapt to requests through prompts. A fine-tuned model like MoPHES was trained on tens of thousands of real therapeutic dialogues and has internalized patterns of professional support: active listening, emotional validation, and session structure. For more on the capabilities and risks of LLMs in therapy, see the article ChatGPT as a Therapist: Opportunities and Risks.

What is computational psychiatry and how does it relate to AI therapy? Computational psychiatry uses mathematical models to understand mental disorders. AI therapy is one of its practical applications: models trained on clinical data apply these principles to support people in real time.