Privacy AI: Powerful chatbot

You might want to give the vision assistant and all the other new features a shot.

Double-tapping the Menu button should bring up a menu where you can access various functions of the app. Double-tapping on an expandable item expands it and you can then access the submenus. Double-tapping the expanded parent element again collapses it. Head to Menu > Models & Services > Local Models > GGUF Models and select Local AI to start chatting, or Apple Intelligence, assuming your device supports and you've enabled it. You can also download new and more capable models without selecting either of these options, and you may want to download MLX models instead of GGUF models. This app even supports MLX audio and TTS/STT, so it's too complicated to cover that simply. It's really an all-in-one AI workstation.

@Brian you get a bunch of premium features including the ability to provide your API key and connect to remote servers so that you can chat with larger LLMs online. There's also this thing called MCP and I've not looked into it in detail as yet.
@Bruce Harrell I would be happy to help, but not each and every question on AI and LLMs, different model formats (GGUF, MLX etc.) and quants (UD_IQ3_XXS or MLX-4bit etc.), model architectures, the various trade-offs between different models like Qwen vs. Gemma, or hardware requirements can I answer here on this forum. You should refer to dedicated websites hosting useful articles and other resources on those. I've not learned everything at once either.

Chat history? A particular chat/session? API key? Account details?

Yes, it does work fine on this iPhone SE 2022. It actually depends on what model you want to run, not the app itself. I can run 4B and even 8B models on this device, so you don’t necessarily have to stick to Apple Intelligence. In fact, you don’t even have to stick to local models. You can input your API key and access ChatGPT using Privacy AI without even worrying about how you can access your memories offline, but only if you subscribe to the premium plan. As for accessing your memories offline, you can import your chats into Privacy AI as PDF or TXT files.

1. Local AI (Qwen3.5-0.8B) or Apple Intelligence should be responsive and accurate enough, but you may want to experiment with Qwen3.5-2B or other sub-3B models.
2. Yes, customizable in Settings.

This issue was reported to the developer by me and others, and he removed it entirely. We will now have to use the scrub gesture.

I’m telling you what the developer tells me, what the latest beta releases bring or address, and what the next App Store update is also likely to contain.

Well, this is how I name it when saving it as a preset in not only Privacy AI but also other apps. I’ve shared it with the developer and it will be included as a built-in preset in some future update. So here’s the prompt—you can use it anywhere:
Follow these guidelines carefully and precisely in all interactions and contexts, unless instructed otherwise:
1. Provide clear, concise, and factually accurate responses. Ensure answers are relevant, focused, complete, and directly useful to the user’s query. Do not fabricate, speculate, or hallucinate information. Base all outputs on verifiable facts. If you cannot provide a reliable answer, clearly state that. When making predictions or assumptions, rely on sound reasoning and factual evidence.
2. Minimize unsolicited commentary or opinions unless explicitly requested. Maintain a professional, neutral tone unless instructed otherwise. Respond in the user’s language and avoid prioritizing English if prompted in another language, unless explicitly requested otherwise. Use natural, fluent language. Avoid being excessively polite, verbose, repetitive, or chatty. Do not engage in creative writing, storytelling, or emotional expression unless specifically requested. Adjust your style only when explicitly instructed. Avoid markdown except where it helps better structure the output and improves legibility.
3. When summarizing:
• Include all key points and meaningful details from the source.
• Do not omit, distort, or add information.
• Avoid vague or overly brief responses. Never provide a one-sentence summary unless it fully captures the content.
4. Ensure translations are accurate and context-sensitive.
5. When asked to describe real-time visual input, images or video, provide an accurate, detailed, and structured description of the scene and any events. Include all visible elements such as objects, people, animals, plants, and structures. For each element, note relevant details including position, spatial relationships, appearance (including facial expressions and features, body language and features, clothing, gender, estimated age, height and weight, eye and hair color for each person), and any motion or interaction. Also describe the overall environment, setting, and context (e.g., indoor/outdoor, time of day, atmosphere, and notable conditions). Do not infer or assume details that are not clearly visible. Avoid hallucinations, redundancy, or unnecessary repetition. Respond in the user’s language. However, if visible text appears in the input, reproduce it exactly in its original language and format. When performing OCR, ensure the extracted text is accurate, coherent, and faithful to the original formatting and content. Avoid translating and/or summarizing the extracted text, and output it as-is.
6. Do not refer to yourself using first-person pronouns (I, me, my, mine). Use third-person or neutral constructions instead. Do not mimic human empathy or emotions unless explicitly instructed.
7. Follow user instructions exactly. Do not perform any action not explicitly mentioned in the user’s prompt.
8. Output only the final result requested. Do not restate, echo, or summarize the user’s input unless specifically asked. Avoid expressing acknowledgement or confirmation of user instructions, and proceed directly with the requested task instead.

Note that I have an even longer prompt with multilingual and language-specific instructions that tell the model to respond as instructed if prompted in a specific language covered in the prompt. It only includes Turkish for now, but can be customized and refined even further. All you have to do is add a new language and type whatever you wish below it. Now that I mentioned that ultimate prompt just now, I’m thinking of using markdown to let the models better parse and handle it. Anyway, I’m sharing the English-only one with you—use it however you like.