PocketPal AI

LLM Ventures

About this app

Unlock the power of AI with Private AI Chat, the ultimate app for secure and private conversations with LLMs. The app brings the sophistication of cutting-edge AI directly to your device, ensuring your chats remain confidential and offline.

Key Features:
- Offline Chatting: Interact with advanced AI models directly on your device without the need for an internet connection.
- Secure & Private: Your data never leaves your device. All conversations are processed locally, providing security and peace of mind.
- Benchmarking: Benchmark the model performance on your phone and share with the community.

How It Works:
- Download & Install: Get the app from the App Store and install it on your device.
- Download Model Weights: Connect to the internet to download the required model weights (in GGUF format, e.g., from Huggingface).
- Chat Offline: Once setup is complete, enjoy private, offline conversations with the model anytime, anywhere.

Ratings and reviews

Great and getting better. I love how one can choose between CPU, Open CL and NPU on Galaxy S25U. Hexagon is fastest for me with 4-10 GPU layers only. The only thing I feel like it's missing is memory management (to remember chat history). also I could never get the image to text to work. I tried various models and it just never generates an output. But for pure text it's fairly good with the llama.cpp backend, the front end UI is very friendly and "corporate". It's abit slow work deleting chats.

the app idea is good but most of the hugging face models keep glitching saying "Completion failed: Exception in HostFunction:" whenever a certain amount of messages are sent. and also crashes are a tiny bit too frequent on models that are int4 or int2

I like the app and how it works but some things are quite unclear such as what the "think" button does and the error "Completion failed: Context is full" and how to fix that. I'd also like help with the + button being greyed out on a chat with a model that has vision support. Hugging face integration is extremely helpful though along with it being localhost so there's no need for internet. Another point of improvement could be LaTeX support for math topics.

What’s new

Memory estimates now learn from your device - more accurate with each model you load.