Skip to main content
PromptQuorum
Home/Power Local LLM/PocketPal AI Review (2026): Free, Open-Source On-Device Chat
Mobile & Edge LLMs

PocketPal AI Review (2026): Free, Open-Source On-Device Chat

Β·10 min readΒ·By Hans Kuepper Β· Founder of PromptQuorum, multi-model AI dispatch tool Β· PromptQuorum
Read in:
πŸ‡ΊπŸ‡ΈenπŸ‡©πŸ‡ͺdeπŸ‡«πŸ‡·frπŸ‡―πŸ‡΅jaπŸ‡¨πŸ‡³zhπŸ‡ͺπŸ‡ΈesπŸ‡§πŸ‡·ptπŸ‡ΈπŸ‡¦arπŸ‡°πŸ‡·ko

PocketPal AI is a free, open-source (MIT license) app that runs GGUF language models entirely on your iPhone, iPad, or Android device, with no cloud dependency for chat. Built by developer Asghar Ghorbani and hosted on GitHub (over 8,000 stars at review time), it ships with quick-download access to models like Gemma 2, Qwen, Phi, and Danube 3, and can load any other GGUF file you source from Hugging Face. The core app costs nothing; an optional in-app marketplace called PalsHub sells premium assistant presets, but chatting with a downloaded model is free.

PocketPal AI is a free, MIT-licensed, open-source app for iPhone, iPad, and Android that runs GGUF language models directly on the device, with no server round-trip and no account required to chat. It was built by Asghar Ghorbani, a developer who started it as a personal project before open-sourcing the code on GitHub. This review covers what the app does today, which models it ships with, its real hardware requirements, its actual pricing (the base app is free; an optional community marketplace called PalsHub sells premium presets), and who should β€” and should not β€” use it instead of a commercial alternative.

PocketPal AI Review (2026): Free, Open-Source On-Device Chat

Key Takeaways

  • Free and open source (MIT license); source code and issue tracker are public on GitHub.
  • Built by Asghar Ghorbani as a personal project, then open-sourced once it gained traction.
  • Runs GGUF models via llama.cpp entirely on-device β€” no account, no server round-trip to chat.
  • Ships with quick-download access to Gemma 2, Qwen, Phi, and Danube 3; any other GGUF model from Hugging Face can be loaded manually.
  • Available on the App Store (iPhone/iPad) and Google Play (Android); no Mac, Windows, or Linux release.
  • The base app has no subscription; an optional PalsHub marketplace sells premium assistant presets via in-app purchase.

πŸ“ In One Sentence

PocketPal AI is a free, MIT-licensed, open-source app for iPhone, iPad, and Android that downloads and runs GGUF language models directly on the device, built and maintained by developer Asghar Ghorbani.

πŸ’¬ In Plain Terms

It is an app you install to chat with an AI model that runs on your own phone instead of a company server β€” the code is public on GitHub, the app itself costs nothing, and you pick which model to download rather than being locked to one provider.

πŸ“ŒNote: This review is based on the public GitHub repository, the project's in-repo documentation, and the App Store and Google Play listings, checked in September 2026. PromptQuorum has not conducted independent hands-on benchmarking of PocketPal AI for this review β€” performance figures below are general local-LLM hardware guidance, not measurements specific to this app.

What Is PocketPal AI?

**PocketPal AI is a mobile app that downloads open-weight language models in the GGUF format and runs them locally on an iPhone, iPad, or Android device, using llama.cpp as the inference engine.** The app itself is built with React Native and is MIT-licensed, meaning the full source code is public and free to inspect, fork, or modify.

Unlike apps that lock users into a fixed, curated model list, PocketPal AI lets you pick from a set of models the app can fetch in one tap, or import any GGUF file you find on Hugging Face β€” including gated models you have accepted the license terms for. That flexibility is the app's defining trait: it behaves more like a general-purpose GGUF runtime with a chat UI than a single-vendor product.

Who Built PocketPal AI?

PocketPal AI was created by Asghar Ghorbani, who started it as a personal side project to run small language models on his own phone before open-sourcing it once it attracted outside interest. Ghorbani wrote about the project's origin and motivation in a Medium post, "PocketPal AI: Tiny LLMs in the Pocket", describing the goal as putting the power of small language models directly in his pocket.

The project is maintained in the developer's spare time rather than by a funded company or dedicated team β€” the repository's own setup documentation notes this directly, which is worth knowing before expecting enterprise-grade support response times. The GitHub repository (a-ghorbani/pocketpal-ai) has accumulated over 8,000 stars and more than 850 forks at review time, and it continues to receive regular releases covering new model support, UI changes, and feature additions such as internet search integration, chat pinning, and export to Markdown.

  • Developer: Asghar Ghorbani (GitHub: a-ghorbani), working independently rather than as a company.
  • Started as a personal project to run small language models on-device; open-sourced after gaining public interest.
  • License: MIT β€” permissive open source, free to use, fork, and modify.
  • Over 8,000 GitHub stars and more than 850 forks at review time, with an active issue tracker and pull request history.
  • Development pace: frequent point releases adding model-runtime updates, new UI features, and expanded language support, maintained on a volunteer basis.

How to Get Started with PocketPal AI

**Getting a working local chat in PocketPal AI takes four steps after installing the app, per the project's own getting-started documentation.** There is no account creation or sign-in step β€” the entire flow happens on-device.

  1. 1
    Install the app
    Why it matters: Download PocketPal AI from the [App Store](https://apps.apple.com/us/app/pocketpal-ai/id6502579498) on iPhone/iPad or [Google Play](https://play.google.com/store/apps/details?id=com.pocketpalai) on Android. No account or email sign-up is required to open the app.
  2. 2
    Open the Models page
    Why it matters: Navigate to the Models tab from the app menu to see the list of quick-download models (such as Gemma 2, Qwen, Phi, and Danube 3) and the option to import a custom GGUF file.
  3. 3
    Download a model
    Why it matters: Pick a model sized for your device's RAM (see the hardware section below) and download it. Larger, more capable models take longer to download and use more on-device storage.
  4. 4
    Load the model and start chatting
    Why it matters: Tap "Load" to bring the downloaded model into memory, then switch to the Chat page. Once loaded, all inference happens on-device β€” PocketPal AI's "Auto Offload/Load" feature can automatically unload the model when the app is backgrounded to save memory, then reload it when you return.

Which Models Does PocketPal AI Support?

Gemma 2

Access method:
One-tap quick download in-app
Typical use case:
General chat; Google's open-weight model family

Qwen

Access method:
One-tap quick download in-app
Typical use case:
General chat and reasoning; Alibaba's open-weight model family

Phi

Access method:
One-tap quick download in-app
Typical use case:
Compact reasoning; Microsoft's small-model family

Danube 3

Access method:
One-tap quick download in-app
Typical use case:
Lightweight chat; H2O.ai's small-model family

Any other GGUF model

Access method:
Manual import via Hugging Face Hub search inside the app, including gated models after accepting license terms
Typical use case:
Custom or niche models not in the quick-download list

Is PocketPal AI Available on iPhone and Android?

iPhone / iPad

Availability:
Available on the App Store
Notes:
Uses Apple's Metal GPU API for acceleration by default; this can be disabled in-app if it causes stability issues on a given device.

Android

Availability:
Available on Google Play
Notes:
Recent releases restored a Qualcomm Hexagon NPU acceleration backend for compatible chipsets, in addition to standard CPU inference.

Mac / Windows / Linux

Availability:
No desktop release found at review time
Notes:
PocketPal AI is a mobile-only app. Desktop users who want a local GGUF chat client should look at Ollama, LM Studio, or Jan AI instead.

How Much Does PocketPal AI Cost?

PocketPal AI is free to download and free to use for chatting with a locally downloaded model β€” there is no subscription, and the source code is public under the MIT license. The only paid component found in the app is PalsHub, an in-app marketplace where community members can publish and sell premium "Pals" (customizable assistant presets); PalsHub in-app purchases are available on the iOS and Android app depending on region.

Because the app and its model-running functionality are open source and free, the only ongoing cost to the user is the storage space and bandwidth needed to download GGUF model files β€” a 3–4B parameter quantized model is typically a few gigabytes, and larger models are proportionally bigger.

  • Base app: Free, no subscription, MIT-licensed open source.
  • Chatting with downloaded models: Free β€” no per-message or per-token charge, since inference runs on your own device.
  • PalsHub marketplace: Optional in-app purchases for premium community-made assistant presets; the base chat functionality does not require this.
  • Storage cost: Not monetary, but real β€” each downloaded GGUF model consumes device storage, typically several gigabytes per model.

What Features Does PocketPal AI Have?

Beyond basic local chat, PocketPal AI has grown a set of features common to more mature local-AI apps, added incrementally across its release history. These include:

  • Pals and PalsHub. "Pals" are customizable assistant presets (system prompt, model, and behavior bundled together); PalsHub is a marketplace where users can share or sell their own Pals.
  • Talents (tool-calling). The app supports an agent-style flow where a model can invoke defined tools mid-conversation rather than only generating plain text.
  • Internet search in chat. Recent releases let a chat search the web using a user-supplied API key from providers such as Brave, Tavily, or Exa β€” this is optional and requires the user's own key, not a bundled search service.
  • Text-to-speech output. The app can read responses aloud, with support expanding across dozens of languages in recent releases.
  • Chat pinning and Markdown export. Conversations can be pinned for quick access and exported as Markdown files.
  • Speculative decoding (experimental). A smaller "draft" model can be paired with a larger model to speed up generation β€” flagged by the project itself as an experimental feature, not a stable guarantee.
  • Benchmarking tools. The app includes a way to benchmark model performance on the current device, useful for comparing how different quantization levels perform on your specific phone.

What Hardware Do You Need to Run PocketPal AI?

The model you pick, not the app itself, sets your real hardware floor β€” as a general rule for GGUF models on phones, a quantized 2–4B parameter model needs roughly 3–4 GB of free RAM, while a 7–8B parameter model needs 6–8 GB or more. This is standard local-LLM sizing guidance, not a PocketPal-specific benchmark; PromptQuorum has not independently benchmarked PocketPal AI's exact memory usage per model.

On iPhone and iPad, the app uses Apple's Metal GPU API by default, which generally improves inference speed over CPU-only execution on Apple Silicon devices. On Android, recent releases added support for a Qualcomm Hexagon NPU acceleration backend on compatible chipsets; devices without a supported NPU fall back to CPU inference, which is markedly slower for larger models.

  • Small models (2–4B parameters, quantized): usable on most phones from the last few years with 4 GB+ of free RAM.
  • Mid-size models (7–8B parameters, quantized): recommended device RAM of 8 GB or more for acceptable responsiveness.
  • Storage: budget several gigabytes of free device storage per downloaded model file.
  • Faster inference: an Apple Silicon iPhone/iPad (Metal acceleration) or a recent Android device with a supported Qualcomm Hexagon NPU will outperform older or budget hardware running CPU-only.

Who Should Use PocketPal AI

  • Privacy-conscious users who want zero subscription cost. If the goal is running an open-weight model with no ongoing fee and no account, PocketPal AI's MIT-licensed, free core app is a direct fit.
  • Users who want to try models beyond a fixed curated list. Because the app can load any GGUF file from Hugging Face β€” not just its quick-download shortlist β€” it suits people who already have a specific model in mind.
  • Developers and technically curious users comfortable with open-source projects. The public GitHub repository, issue tracker, and frequent releases suit users willing to read release notes or file an issue rather than expect dedicated customer support.
  • Users who want to inspect or modify the code themselves. The MIT license permits forking and modification, which matters to anyone who wants to audit exactly what the app does before trusting it with private conversations.

Who Should Not Use PocketPal AI

  • Users who want a fully guided, zero-configuration experience. Picking the right model size for your device and managing downloads yourself is part of the workflow; users who want the app to make that decision for them may prefer a more curated app like Loci.
  • Users who need Mac, Windows, or Linux support. PocketPal AI is mobile-only (iPhone/iPad and Android); desktop users should look at Ollama, LM Studio, or Jan AI instead.
  • Users who want the most demanding models (70B+ parameters). Phone hardware, even flagship devices, cannot practically run models at that scale locally β€” those workloads need a desktop, workstation, or cloud API regardless of which mobile app is used.
  • Users who expect dedicated commercial support. PocketPal AI is maintained by one developer in spare time, per the project's own documentation; users who need guaranteed response times or enterprise support agreements should look at a commercially backed alternative.
  • Users who want built-in cloud sync across devices. Because inference and chat history are local to each device by design, PocketPal AI does not offer a built-in account-based sync between an iPhone and an Android device the way a cloud-backed chat app would.

PocketPal AI vs. Alternatives

PocketPal AI

License / cost:
Free, open source (MIT)
Platforms:
iPhone/iPad, Android
Model flexibility:
Quick-download shortlist (Gemma 2, Qwen, Phi, Danube 3) plus any GGUF file from Hugging Face
Articles about PocketPal AI (10)

+6 more not shown

Loci

License / cost:
Free tier plus paid tiers; closed source
Platforms:
iPhone/iPad/Android/Mac/Windows (5 platforms)
Model flexibility:
Curated library (~10 models); cannot import custom GGUF files

Private LLM

License / cost:
Paid, one-time purchase; closed source
Platforms:
iPhone/iPad/Mac (Apple only)
Model flexibility:
140+ models, with OmniQuant and GPTQ quantization formats
Articles about Private LLM (10)

+4 more not shown

Enclave AI

License / cost:
Free with paid tier; closed source
Platforms:
iPhone/iPad/Mac (Apple only)
Model flexibility:
Curated on-device model list; Siri and Shortcuts integration

Locally AI

License / cost:
Free; closed source
Platforms:
iPhone/iPad
Model flexibility:
Curated model list, per its App Store listing

ChatterUI

License / cost:
Free, open source (AGPL-3.0)
Platforms:
Android (build from source)
Model flexibility:
Runs GGUF files on-device via llama.cpp, or connects to remote APIs; Character Card v2 support

RikkaHub

License / cost:
Free, open source
Platforms:
Android
Model flexibility:
Multi-provider client; local on-device inference via a LiteRT provider for models like Gemma and Qwen
Articles about RikkaHub (1)

Also mentioned in:

MLC Chat (MLC-LLM)

License / cost:
Free, open source (Apache 2.0)
Platforms:
iPhone/iPad, Android
Model flexibility:
Runs models compiled for the MLC runtime rather than raw GGUF files

LLM Farm, a formerly popular open-source iOS GGUF chat app, was delisted from the App Store and marked "temporarily unavailable" on its own GitHub repository at review time β€” it is omitted from active recommendations above for that reason. Layla, another on-device assistant app, reportedly had its free "Layla Lite" tier removed from Google Play; check layla-network.ai directly for current availability before assuming it is installable.

Frequently Asked Questions

Is PocketPal AI free?

Yes. PocketPal AI is free to download and free to use for chatting with a downloaded model, and the app is open source under the MIT license. An optional in-app marketplace, PalsHub, sells premium community-made assistant presets, but the core chat functionality has no subscription or paywall.

Who developed PocketPal AI?

PocketPal AI was created by Asghar Ghorbani, who started it as a personal project to run small language models on his own phone before open-sourcing the code once it gained outside interest. It is maintained independently rather than by a company.

Is PocketPal AI open source?

Yes, the full source code is published on GitHub under the MIT license, a permissive license that allows free use, modification, and redistribution.

Which models can PocketPal AI run?

PocketPal AI ships with one-tap downloads for Gemma 2, Qwen, Phi, and Danube 3, and can additionally load any other GGUF-format model you find on Hugging Face, including gated models after you accept their license terms.

Does PocketPal AI work on iPhone and Android?

Yes. PocketPal AI is available on the App Store for iPhone and iPad, and on Google Play for Android. There is no Mac, Windows, or Linux version.

How much RAM does PocketPal AI need?

This depends on the model you choose, not a fixed app requirement. As general guidance for quantized GGUF models on phones, 2–4B parameter models typically need 3–4 GB of free RAM, and 7–8B parameter models typically need 8 GB or more. Larger devices with more RAM can run larger, more capable models.

Does PocketPal AI require an internet connection?

No internet connection is required to chat once a model is downloaded β€” inference runs entirely on-device via llama.cpp. An internet connection is needed to download models initially, and optionally to use the app's internet-search feature, which requires the user's own API key from a provider such as Brave, Tavily, or Exa.

Is PocketPal AI safe for private conversations?

Because inference happens locally and the source code is public and auditable under the MIT license, PocketPal AI does not need to send chat content to a server to function. This review has not conducted an independent security audit of the app; users who need that level of assurance should review the public repository or its issue tracker directly.

How does PocketPal AI compare to Private LLM or Loci?

PocketPal AI is free and open source with unrestricted GGUF model loading, while Private LLM is a paid, closed-source Apple-only app with a larger built-in model catalog (140+ models), and Loci is a closed-source app spanning five platforms with a smaller, more curated model library. Choose PocketPal AI if cost and openness matter most; choose Private LLM or Loci if a more guided, polished experience across more platforms matters more than being free or open source. See the comparison table above for details.

Can I run large models like 70B parameter models on PocketPal AI?

No. Phone hardware, including current flagship devices, does not have enough RAM to practically run models at that scale locally. PocketPal AI is built for small and mid-size models (roughly up to 8B parameters); users who need larger models should use a desktop workstation or a cloud API instead.

Verdict

PocketPal AI earns its place among mobile local-AI apps by being genuinely free, genuinely open source, and genuinely flexible about which model it runs β€” three things that are each individually rare, and rarer still combined, in this category. It is not the most polished app in this space: users manage their own model downloads, the interface trades some guidance for flexibility, and the project is maintained by one developer in spare time rather than a company with a support team. Choose PocketPal AI if you want to run an open-weight model on your own phone with no subscription and no vendor lock-in, and you are comfortable picking a model sized for your device yourself. Choose a paid, curated app like Private LLM or Loci instead if you would rather trade some cost and openness for a more guided setup and dedicated support across more platforms.

Sources

← Back to Power Local LLM