9 Offline AI Apps for Android and iPhone, Compared on the Record
Nine local AI apps for Android and iPhone, compared on facts read off their own store listings and repositories. Engine, price, licence, install base and last update, with everything we could not verify called out.
TokForge is one of the nine apps below and this is the TokForge website. We have not ranked the list, we have not given ourselves a trophy, and we have not written a paragraph about how our app came out on top. What we have done is put every app's verifiable facts in one table and told you where each fact came from, including the ones that are unflattering to us.
How this comparison was made, and what it is not
Every cell in the table was read on 2026-07-27 from the app's own Google Play listing, its App Store entry through Apple's public lookup API, or its public source repository. Nothing was copied from another roundup, which matters more than it sounds: several widely-shared articles still list a price that changed, an engine that was never used, and an app whose last release was in 2024.
We did not run a cross-app speed test. That is the part every other roundup leads with, so here is why we left it out. To make a tokens-per-second comparison across nine apps mean anything, you have to hold the model, the quantization, the context length, the thermal state and the handset constant across all of them, and most of these apps do not even offer the same model. A number produced by starting a stopwatch on three phones is not evidence, and we would rather publish a smaller honest thing than a large confident one. Where we do publish speed, it is per-device and dated on our public leaderboard, and it is submissions from real handsets rather than a number we generated ourselves.
The table
| App | Platforms | Price | Source | Inference engine | Public speed board | Play installs | Last update |
|---|---|---|---|---|---|---|---|
| Google AI Edge Gallery | Android, iOS, macOS | Free | Apache-2.0 | LiteRT / LiteRT-LM | No | 1M+ | 2026-07-22 |
| PocketPal AI | Android, iOS | Free, with in-app purchases | MIT | llama.cpp (GGUF) | Yes | 1M+ | 2026-07-11 |
| Layla | Android, iOS | Paid, iOS upfront plus in-app purchases | No public repo found | llama.cpp, LiteRT-LM, ExecuTorch, QNN | No | 10K+ | 2026-07-10 |
| Private LLM | iOS, macOS, Android beta | $4.99 once | No public repo found | mlc-llm | No | - | 2026-07-09 |
| ChatterUI | Android | Free | AGPL-3.0 | llama.cpp (GGUF) | No | Not on Play | 2026-04-24 |
| LM Playground | Android | Free | MIT | llama.cpp (GGUF) | No | 50K+ | 2026-07-13 |
| Enclave AI | iOS, macOS | Free, with a paid tier | No public repo found | llama.cpp | No | - | 2026-04-19 |
| MLC Chat | iOS, Android APK | Free | Apache-2.0 | MLC / TVM | No | No Play listing found | 2024-09-27 |
| TokForge (ours) | Android, iPhone beta | Free | No public repo | llama.cpp (GGUF) and MNN | Yes | 1K+ | 2026-07 |
TokForge is free. Android is on Google Play open testing; iPhone and iPad are in open TestFlight beta. No account, and it works with the network off once a model is downloaded.
Reading on a computer? See real measured speeds on the phone-speed leaderboard.
The apps
Google AI Edge Gallery
Google's own demo app and the largest installed base in the category. It is a gallery for trying models rather than a chat app you would live in, and it benchmarks locally without publishing anything. Now on iOS and macOS as well, which several roundups still get wrong. No star rating is displayed on Play.
PocketPal AI
The other 1M+ app, and the only competitor here that publishes a public speed board. Its board is a pairwise ranking rather than a raw table. Open source under MIT. If you want the safest recommendation for someone who has never run a local model, this is it.
Layla
Genuinely multi-backend, which is rarer than the marketing in this category suggests. Roleplay and character focused. The paid model is the main thing separating it from the free options rather than any capability gap.
Private LLM
Apple-first and the most polished of the paid options. Worth correcting two things that get repeated: it is $4.99, not $9.99, and it is not a llama.cpp wrapper. It also runs by far the largest set of per-model web pages in this category, which is how most people find it.
ChatterUI
Distributed through GitHub releases, not a store, so there is no install count to quote. Popular with the roleplay crowd and well regarded on Reddit, which is where it is discovered.
LM Playground
Small, actively maintained, does the one job without ceremony.
Enclave AI
Apple only, and the highest App Store rating in this set. Nothing for Android users here.
MLC Chat
Read that update date again. The MLC engine repository is very much alive, but the shipped iOS app has not been updated since September 2024. Several 2026 roundups still rank it as a current app, and at least one calls it the fastest Android option. It is a demo for an actively developed engine, not current consumer software.
TokForge
This is our app, so treat the rest of this entry as disclosure rather than a review. It runs both GGUF and MNN, which no other app in this list does, and it publishes a raw public speed board with the device, chip, engine, app build and date on every row. It is much smaller than the incumbents at 1K+ installs. The Android listing serves an open testing track rather than a production release, and the iPhone build is TestFlight only and is not on the App Store. Neither store displays a star rating for it, so there is not one to quote.
Five things the popular roundups get wrong
MLC Chat is listed as a current app. Its shipped iOS build dates from September 2024. The engine behind it is actively developed, which is probably why the confusion persists, but the app you would install is nearly two years old.
Private LLM's price and engine. Widely quoted at $9.99 and described as a llama.cpp or MLX wrapper. It is $4.99 and it runs on mlc-llm.
Google AI Edge Gallery described as Android only. It has shipped on iOS and macOS.
Star ratings compared across stores. Play review counts and App Store rating counts are different quantities and putting them in the same column implies a comparison that is not there.
Speed numbers with no fence around them. A tokens-per-second figure without the handset, the model, the quantization and the date attached is decoration. This is the failure mode we audited our own site for, and we removed a table for exactly this reason.
What to actually pick
If you have never run a model locally, install PocketPal AI or Google AI Edge Gallery. Both are free, open source, maintained, and large enough that any problem you hit has been discussed somewhere.
If you already know what you want, the question is narrower than "which app is best". Character roleplay with imported cards points at Layla or ChatterUI. Apple hardware and a willingness to pay once points at Private LLM. A GGUF collection you already own points at any llama.cpp app. Wanting MNN bundles as well as GGUF, or wanting to see what your specific handset records against everyone else's, points at ours, and you now know exactly how big we are and what track the listing serves.
Whatever you install, check it in airplane mode on day one. That is a thirty-second test and it settles the only claim in this category that really matters.
Common questions
What is the best offline AI app for Android?
For most people arriving at this question the answer is PocketPal AI or Google AI Edge Gallery, because they are free, open source, well maintained and have the largest installed base by a wide margin. If you want a specific capability, the useful question is narrower: character roleplay points at Layla or ChatterUI, running MNN bundles as well as GGUF points at TokForge, and Apple hardware points at Private LLM or Enclave AI.
Do these apps really work offline?
The ones listed here run the model on the handset, so once a model is downloaded they work with the radios off. The download itself needs a network. Be careful with apps advertised as offline that turn out to call a hosted API for anything beyond a small local model; the way to check is to put the phone in airplane mode and see whether it still answers.
Which local AI apps are open source?
As of 2026-07-27: Google AI Edge Gallery (Apache-2.0), PocketPal AI (MIT), MLC Chat (Apache-2.0), ChatterUI (AGPL-3.0) and LM Playground (MIT). Layla, Private LLM, Enclave AI and TokForge do not publish source. Note that finding no public repository is not proof that something is proprietary, it just means we could not verify a licence.
How much RAM do I need to run a local AI model on my phone?
4 GB is the practical floor and it limits you to models under about 1 GB. 8 GB makes the 4B class usable, which is where most people settle. 12 GB is where 8B and 9B models stop feeling like an experiment. Total RAM is not the whole story, because the operating system takes a large share and a backgrounded app gets less again. Our phone RAM calculator works through it with the error bars stated.
Which of these apps publish speed benchmarks?
Two: PocketPal AI and TokForge. Google AI Edge Gallery benchmarks on your device but does not publish the results anywhere. Everyone else publishes nothing. If a roundup quotes tokens per second for six apps, ask which handset, which model, which quantization and which date, because those four numbers are what make the figure mean anything.
Is there an app that runs both GGUF and MNN?
In this set, TokForge, which is ours. Alibaba's own MNN Chat app runs MNN bundles but will not load GGUF files, so it is not an alternative if you have a GGUF collection. Everyone else in the category is llama.cpp underneath, with the exceptions of Private LLM on mlc-llm, MLC Chat on TVM and Google on LiteRT.