Clinically Intelligent is written by Juwon Akinyande in a personal capacity and is not affiliated with, endorsed by, or representative of Barts Health NHS Trust or any other NHS organisation. Content is for general educational purposes only. It does not constitute clinical, legal, or information governance advice. Before applying any guidance to your own practice, consult your Trust information governance lead, your Caldicott Guardian, your line manager, and your professional body.
One of the things I have been learning and keeping an eye on is open source models. These are language models that anyone can use without needing to connect to a company’s servers. The best thing about them is that your data stays on your phone.
One of the biggest questions people ask about consumer AI tools is where your information is stored. With open source models running locally, that question does not apply. Nothing leaves your device. I will go into more detail about open source in later issues once I have learned more. For now, this issue is about the specific tool I have started using. PocketPal.
What PocketPal is.
PocketPal is a free app created by an independent developer and hosted on GitHub at https://github.com/a-ghorbani/pocketpal-ai. It runs open source AI models directly on your phone, which means it works without an internet connection. Nothing you type is sent anywhere. Everything stays on your device.
Why open source matters.
Open source means the code and the model are public. Anyone can inspect what the tool does. Anyone can raise issues, suggest fixes, or fork the project and build on it. If the developer stops working on PocketPal tomorrow, someone else can pick it up because the code is available.
That is a different model from consumer AI. With Copilot or ChatGPT you use what the company decides to build, at the price they set, for as long as they choose to support it. With open source, the tool belongs to the community that uses it.
That does not mean open source is automatically safer or better. It means the tool is visible and the community can hold it accountable. For a lot of people that visibility matters.
Setting it up.
The setup is fairly simple. You download the app from the App Store. You do not need an account. Once you open it, it asks what you plan to use it for. Once you pick a use case, it asks you to pick a model. From what I can tell it offers the same selection of models regardless of the use case, and depending on how much space you have on your phone, you can download whichever one you want.
I chose the best model available at the time of writing. Qwen3.5 4B at 2.5 GB. It took about five minutes to download and be ready to use.
The interface.
PocketPal calls the different assistants pals. Each pal is designed for a specific job. Pip is described as a friendly general purpose pal that runs entirely on your phone. Muse is a creative writing partner for drafting, polishing and finding rhythm. Lookie is a real time video analysis assistant that uses your camera. You pick the pal you want and download the model that powers it.
I could not get Lookie to work. The model would not download for me. I did not spend long trying but it is a feature I will come back to.
Where PocketPal is genuinely useful.
Community work is where signal is least reliable and some hospital buildings are known for their patchy connections. A quick fact check, a service lookup or a patient question that you want to answer in the moment rather than after the visit ends. PocketPal makes those possible in a way consumer AI tools cannot.
Privacy that stays on your device.
There is a common saying that you do not need privacy unless you have something to hide. I don’t agree with that. Everyone is entitled to some form of privacy and PocketPal lets you have it.
In an era where almost everything is on the cloud, having access to a language model that lives entirely on your phone is a different category of tool. It could also be a safer entry point for colleagues who have been apprehensive about AI because they do not know where their queries end up. With PocketPal, they end up on the device and nowhere else.
No credits, no limits.
Another benefit is that PocketPal has no credits or usage limits. You can use it as much as you want without worrying about running out of a monthly allowance. That makes it a useful place to practice. Prompts you have been wanting to try, queries you have been cautious about wasting credits on, ways of testing what different prompt structures produce. All of that becomes low stakes when there is no cost attached.
Information governance.
Running the model locally solves one specific governance question. Where your data goes. Nothing you type into PocketPal leaves your device. That is a real difference from consumer AI tools like ChatGPT, Claude, or Copilot, where every query is sent to a company server.
That difference matters for anyone whose concern about consumer AI has been the data leaving their control. It does not mean local models are cleared for any use. They are not. Composite scenarios still matter. Patient identifiable information still does not go into any AI tool, local or otherwise. The Safety Check from Issue 07 still applies. If your Trust has published guidance on AI use, that guidance still applies here too.
Local models also carry their own risks that the cloud models handle differently. Model quality is slightly lower, which can mean more hallucinations. There is no corporate team behind the tool the way there is with Copilot or ChatGPT. What there is instead is a public GitHub repository where anyone can raise issues, see what has already been reported and follow how the developer is responding. That is a different model of support. Slower but more open. If something goes wrong, you raise it in public and others can see it. If a fix is coming, you can see when.
What running locally changes is one part of the picture. The rest of the discipline stays the same. If anything, it needs slightly more of your own vigilance because there is less institutional infrastructure around the tool.
Limitations.
If you are expecting a model that matches ChatGPT, Claude, or Copilot, this is not that. The model is capable and can handle most queries and tasks that most of us will need, but the larger cloud hosted models will often give you better output.
It is an open source model created by a developer, which means it does not work as smoothly as the tools mentioned above. As one example, I wanted to attach the AI Safety Check to see if I could give the model a composite patient scenario and get it to flag any mistakes. The plus sign that would let me attach documents or images was greyed out, so I could not test it.
Other than a few technical issues like that one, the model works fine.
Where I am with this.
PocketPal is not going to replace ChatGPT or Claude for the work I do most weeks. But for the situations where consumer AI is not the right tool, having a local model on your phone is a real option. Offline settings. Queries you want kept entirely on your device. Prompt practice without cost. Those are legitimate use cases and PocketPal fills them in a way nothing else does.
I will keep exploring what open source models can do and I will report back on what I find.
What you can do this week.
Download PocketPal. Pick a pal. Try one query you would normally ask Claude or ChatGPT. See the difference in the output. That is the fastest way to understand what a local model actually is.
That is all for Issue 15. Next week. What does it mean when we say something is AI.
The AI tools discussed in Clinically Intelligent are consumer products. They have not been independently assessed by the author against DCB0129 or DCB0160 clinical risk management standards, and they may not be approved for clinical use by your employer. Before using any tool described in this newsletter in connection with your clinical practice, you must satisfy yourself that its use is permitted under your Trust information governance policy, your DSP Toolkit obligations, your professional registration requirements, and any applicable contractual terms with your employer. The author accepts no liability for use of any tool or workflow described in this publication. Patient identifiable information must not be entered into any consumer AI tool under any circumstances, irrespective of any guidance contained in this newsletter.