Offline AI roleplay means the model answers from inside your phone, with the radio off. PokiTavern ships the inference engine in the app; you download one small model once, over Wi-Fi, and after that a chat works on a plane, in a tunnel, or with the SIM pulled out.
- App
- PokiTavern 1.1.2, Android 7.0 or newer, 202 MB
- Model
- downloaded in-app, free, 614 MB to 3.66 GB
- Cost
- free; no account, no quota, no subscription
- Data
- local SQLite on the phone
1.What does offline AI roleplay actually mean?
Two very different setups get called offline, and only one of them survives losing signal. Pointing a phone at a PC at home, a tunnel into it, or an Ollama box on your desk keeps your prompts away from any AI vendor, which is genuinely private. The phone still needs a route back to that machine.
Leave the house, board a plane, or lose the tunnel and the second setup goes quiet. PokiTavern puts the model on the handset you are holding, so there is no route to lose. That is the whole difference between private and offline.
2.Do I have to download a model first?
Yes, and it is worth being blunt about that. The 202 MB APK carries the LiteRT inference engine (about 25 MB), four Qualcomm NPU backends (about 44 MB), the Flutter runtime for three ABIs, the SillyTavern frontend it serves (14.9 MB) and 16 MB of seed content. No language model.
On first run you pick one inside the app and it downloads over Wi-Fi, free,
from public Hugging Face mirrors: Qwen3 0.6B at about 614 MB, Gemma 4 E2B at
about 2.59 GB, or Gemma 4 E4B at about 3.66 GB. The URL of any public
.litertlm file works too. Airplane mode is only true after that
download finishes.
3.How good is a model that fits on a phone?
Small, and it reads that way. The largest of the three takes about 3.66 GB on disk, which by cloud standards is nothing. These models hold a character's voice across a handful of exchanges, and the app loads them with a mobile-sized context window rather than their native one, because a full window's key-value cache would not fit in a phone's memory. Long scenes are where they run out first.
So treat the on-device model as the pour that is always available, not as a replacement for a frontier model. When a session deserves better, the same app connects to OpenAI-, Claude-, Gemini- and OpenRouter-compatible endpoints, and to local ones such as Ollama. You bring the key and pay that provider directly.
4.What does it cost, and what leaves the phone?
Nothing, and nothing. There is no subscription, no message quota, no account to create and no telemetry. Characters, chats, presets and world books live in a local SQLite database on the device. With the on-device model there is nobody to send a prompt to, so none is sent.
A token bill only appears if you choose to attach your own API key, and then it appears at your provider, not here. PokiTavern never sits between you and the model you picked.
5.What you need to start
- An Android phone running 7.0 or newer.
- Room for the 202 MB app plus whichever model you pick.
- Wi-Fi for the first model download. Once.
- No Termux, no Node.js, no PC left running at home, no store account.
PokiTavern is not on Google Play yet. The signed APK, its sha256 digest and the release notes are on the download page.
PokiTavern is not affiliated with the SillyTavern project. It embeds a server implementing the SillyTavern 1.18 API and serves that project's official frontend on top.
6.Common questions
Does AI roleplay work with no internet at all?
Yes, once a model has finished downloading. Before that the app has nothing to run on. After it, replies are computed by the phone, so airplane mode, a tunnel or a dead SIM changes nothing about the conversation.
Is the model inside the APK?
No. The 202 MB package holds the LiteRT inference engine, four Qualcomm NPU backends, the Flutter runtime for three ABIs, the SillyTavern frontend and the seed content. Models are downloaded separately in the app, for free.
Do I need an account or an API key?
Neither. The on-device model needs no key, no sign-in, no subscription and no message quota. A key is optional, for the times you would rather route a session to a larger cloud model.
Is a phone model as good as a cloud model?
No. A phone-sized model holds a character's voice over a short exchange and answers quickly. A frontier cloud model is better at long context and complicated scenes. PokiTavern lets you use either one without leaving the app.
Where are my characters and chats stored?
In a SQLite database on the phone. There is no account to attach them to, nothing syncs to a server we run, and the app sends no telemetry.
Can I do this on an iPhone?
Not today. PokiTavern is an Android app and there is no iOS build on the App Store yet. An iPhone on the same Wi-Fi can open the address served by an Android phone, but that is a shared session, not offline chat on the iPhone.