Jan vs PocketPal AI

Jan

• Running open models offline on your own computer
• People who want one desktop app for local and cloud models
• Developers who want a local OpenAI-compatible endpoint

PocketPal AI

• Running open models privately on an iPhone, iPad or Android phone
• Trying many GGUF models from Hugging Face on a mobile device
• Checking how fast a particular phone runs local models

Jan

• Desktop only, on macOS, Windows and Linux. Intel-based Macs are not supported.
• Local models need memory: Jan's Mac guide suggests 8GB of RAM for 3B models, 16GB for 7B models and 32GB for 13B models.
• Cloud models need your own API key from each provider, and those providers see the messages you send them.
• Cowork and the standalone Jan Agent are previews on nightly builds, and the memory feature is marked as coming soon.

PocketPal AI

• It is built for phones and tablets only; the App Store listing marks it as not verified for macOS.
• Which models run, and how fast, depends on the device's memory and chip, so each model's quantization has to be chosen to fit.
• The built-in tools are limited to a calculator, date and time, and HTML rendering.
• Premium Pals are paid extras, and prices differ between the App Store, Google Play and the PalsHub website, which charges in euros.

Jan

Local models

Downloads and runs open models such as Llama, Gemma and Qwen from Hugging Face using llama.cpp or MLX. The Hub shows whether each model fits your hardware.

Cloud providers with your own key

Connects to OpenAI, Anthropic, Gemini, Mistral AI, Groq, OpenRouter and Azure OpenAI, or to any OpenAI- or Anthropic-compatible endpoint, using an API key you supply.

Built-in web search

Web search and page fetch tools are on by default for models that support tool use. They run through a keyless Exa endpoint, so no API key or MCP server is needed.

Projects and assistants

Projects group conversations under shared instructions and files, and custom assistants keep their own instructions for recurring kinds of work.

Local API server

A built-in OpenAI-compatible server at localhost:1337 lets other applications use the models running in Jan, and Claude Code can be set to use them as its backend.

MCP connectors and Cowork preview

MCP servers add tools such as code execution and databases. Cowork, which works through multi-step tasks on files and commands, ships only in nightly builds as a preview.

PocketPal AI

On-device chat with GGUF models

Runs open models such as Gemma, Qwen, Phi and Llama through llama.cpp on the phone's CPU, GPU or, on Qualcomm chips, NPU. Nothing is sent to a server.

Hugging Face downloads

Search, bookmark and download GGUF models from the Hugging Face Hub inside the app, picking a quantization that fits the device. Gated models work with your own access token.

Pals and PalsHub

A Pal is an assistant with its own default model, system prompt and personality, in Assistant or Roleplay form. PalsHub lists Pals made by other people, both free and paid.

On-device text-to-speech

Replies can be read aloud by neural voice engines such as Kokoro that run on the phone, so speech needs no cloud service either.

Built-in tools

Capable Pals can call tools in the middle of a conversation: a calculator, the current date and time, and rendering of HTML that the model writes.

Benchmarking and leaderboard

Measures tokens per second and memory use for a model on your device. Results can optionally be submitted to the public AI Phone Leaderboard.

Jan

• Free$0

  The full desktop app for macOS, Windows and Linux, with local models, cloud provider connections and the local API server.

PocketPal AI

• Free$0

  The full app: offline chat with GGUF models, Hugging Face downloads, text-to-speech, your own Pals, built-in tools and benchmarking.

• Premium Pals$0.99 to $2.99

  per Pal, one-time in-app purchase (US App Store)

  Optional community-made assistants from PalsHub. Google Play's US listing shows $1.39 to $2.79 per item, and the PalsHub website prices paid Pals from €1.49.