Your Phone Can Run Real AI Now And That Changes Apps
Every figure, product claim and quoted description the finished picture puts on screen,
chased to a primary source.
PocketPal AI
Source: https://github.com/a-ghorbani/pocketpal-ai
- Describes itself as “A private AI assistant that runs entirely on your phone.” The
repository elaborates: “Chat with language models, give them a voice, and let them use
tools — all on-device. No account, no cloud, no internet required.”
- Runs GGUF language models (Gemma, Qwen, Phi, Llama and others) fully offline.
- “Hugging Face integration — search and download GGUF models directly from the HF Hub.”
- “Benchmarking — measure tokens/sec and memory.”
- “Hardware acceleration — CPU, GPU (Metal on iOS, OpenCL/Adreno on Android), and NPU
(Qualcomm Hexagon) inference paths.”
- “Text-to-speech — on-device neural TTS with no cloud calls.”
- “Works offline — no connection and no account required.”
- “Private by default — every prompt, response, and document stays on your device.”
- MIT licensed, free and open source, no subscription.
- Built on React Native and TypeScript, using llama.cpp for LLM inference and ONNX Runtime
for text to speech. This is the link the script draws between PocketPal and the desktop
world of quantized GGUF models: it is the same inference engine.
On screen: the GGUF format name, the Hugging Face download step, the CPU / GPU / NPU
acceleration paths, the Qualcomm Hexagon NPU label, and the phrase “no account, no cloud,
no internet required” as an attributed description rather than a channel claim.
Android AICore and Gemini Nano
Source: https://developer.android.com/ai/aicore
- AICore is an Android system service that enables on-device execution of generative AI
foundation models. Quoted: “Gemini Nano runs in Android’s AICore system service, which
leverages device hardware to enable low inference latency and keeps the model up-to-date.”
- The model is “Google’s Gemini Nano foundation model.”
- Stated benefits, which are the three the script names:
- Privacy: “On-device generative AI executes prompts locally, eliminating server calls”
and “keeps sensitive data on the device.”
- Offline: works without a network connection.
- Performance: “low inference latency” and “accelerated inference” through on-device
hardware.
- Also stated, and used in the platform-plumbing argument: “AICore manages the distribution
of Gemini Nano and handles future updates. You don’t need to worry about downloading or
updating large models over the network.”
- Cost: “Significantly reduces the cost of using these large models in your app.”
Not on screen: a device support list. The Android documentation does not publish one on
this page, and the narration says “supported Android devices” rather than naming any, so no
device names or model numbers are rendered.
Google AI Edge Gallery
Source: https://github.com/google-ai-edge/gallery
- Describes itself as “A gallery that showcases on-device ML/GenAI use cases and allows
people to try and use models locally.”
- “Running the world’s most powerful open-source Large Language Models (LLMs) on your
mobile device.”
- “100% On-Device Privacy: All model inferences happen directly on your device hardware.”
- Model management: “Easily download models from the list or load your own custom models.
Manage your model library effortlessly and run benchmark tests.”
- Prompt Lab: “A dedicated workspace to test different prompts and single-turn use cases
with granular control over model parameters.”
- Mobile Actions: “Unlock offline device controls and automated tasks powered entirely by a
finetune of FunctionGemma 270m.”
- Platforms: Android 12 and up, iOS 17 and up, and a downloadable macOS build.
On screen: the four capabilities the narration lists (run models locally, test prompts,
benchmark performance, mobile actions), and the Android 12+ / iOS 17+ platform floor.
FunctionGemma 270M
Sources:
On screen: the figure 270M, the Gemma 3 lineage, and one of Google’s own example
commands rather than an invented one.
Apple Foundation Models framework
Sources:
On screen: the ~3B on-device parameter figure, the @Generable guided-generation
mechanism, tool calling, and the “won’t increase your app size” claim as an attributed one.
Private Cloud Compute
Sources:
Not on screen: the on-device versus Private Cloud Compute context window figures. Those
were only found in secondary summaries, not on an Apple page that could be read directly, so
no context-window number is rendered anywhere in the video. See “Not verified” below.
MLC LLM and MLC Chat
Sources:
On screen: the platform and backend matrix, Metal and OpenCL named against iOS and
Android, and the two App Store sentences quoted and attributed.
GGUF, llama.cpp and quantization
Source: https://github.com/a-ghorbani/pocketpal-ai (states llama.cpp as the inference
engine and GGUF as the model format it consumes)
- GGUF is the model file format used by llama.cpp, which is the same format desktop users
already run locally. PocketPal consuming GGUF from the Hugging Face Hub is what makes the
script’s claim true that a phone can use “the same broad world of quantized models people
already use with llama.cpp on desktops”.
On screen: the GGUF extension and the llama.cpp name, as the shared format link between
the desktop and the phone. No quantization bit-depth figures are rendered, because the
script names none and none was chased.
Not verified
Claims the narration makes that could not be chased to a primary source, and which are
therefore not rendered as figures anywhere in the picture:
- “Three years ago, the normal advice was simple. If you wanted serious AI, you needed a
data center.” This is a characterisation of how the field was discussed rather than a
sourceable fact, and the shot treats it as a remembered claim rather than a cited one.
- The relative battery cost, heat and throttling behaviour of on-device inference. Both
Apple and Google state that on-device inference uses device resources, but no published
primary figure for battery drain per token was found, so no percentage, milliamp-hour or
minutes-of-battery number appears on screen.
- The comparative economics of local versus cloud inference for an app developer. The
direction is supported by Apple stating on-device inference is “free of cost” and Google
stating AICore “significantly reduces the cost”, but no per-request price is rendered.
- On-device versus Private Cloud Compute context window sizes. Only found in secondary
summaries, so no context figure is shown.