Can You Use ChatGPT Offline?
No, and no setting will change that. But you can run a different model entirely on your phone. Here is the honest version of what that gets you.
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 direct answer: no. ChatGPT itself cannot work offline. The app on your phone is a thin client. Every message you type is sent to OpenAI's servers, the model runs there, and the reply comes back over the network. Turn the network off and the app has nothing to talk to. The same is true of Gemini, Claude and Copilot. There is no hidden setting, no download, no offline mode.
What you can do is run a different model, one whose weights sit on your phone's storage, entirely on the handset. That is a real thing you can have working in a few minutes, and this page explains what it is actually like rather than overselling it.
Why people want this
- No signal. Planes, trains, basements, hiking, rural areas, travelling without roaming data.
- Privacy. Some conversations should not be transmitted anywhere. On-device means nothing to intercept and no account tied to what you asked.
- No subscription. A local model does not meter you or expire.
- No rate limits or outages. The model is on the device; it does not go down.
The honest trade-off
A model small enough to run on a phone is not GPT-5 and will not pretend to be. It is worth being blunt about where the line falls, because the biggest cause of disappointment here is a page like this overselling it.
| Task | How an on-device model does |
|---|---|
| Drafting, rewriting, summarising, tone changes | Genuinely good. This is the strongest use case. |
| Brainstorming, naming, outlining | Good. Speed matters more than depth here, and local is fast. |
| Translation and language practice | Good on the mid-size models. |
| Character chat and interactive fiction | Good, and a common reason people prefer local. |
| Questions about a document you supply | Good, because the document is the context rather than the model's memory. |
| Code | Workable at small scale. Do not expect a frontier coding model. |
| Hard multi-step reasoning and long-horizon maths | Weak. This is where the size gap really shows. |
| Current events and facts after training cutoff | It does not know them, and it may confidently invent them. Optional web search exists for this, but that requires a connection by definition. |
If someone tells you a phone-sized model matches ChatGPT, they are selling something. What is true is that for a large share of everyday writing and thinking tasks, the gap is much smaller than people expect, and the model answers instantly with nothing leaving the device.
What you need
- An Android phone on 8.0 or later, or an iPhone or iPad on iOS 16 or later.
- Ideally 8 GB of RAM. 6 GB works with smaller models. 4 GB runs the small ones only.
- A few gigabytes of free storage for the model file.
- A connection once, to download the model. After that it runs with the network off.
Getting it working
- Install TokForge. Free on Google Play, and in open TestFlight beta on iPhone and iPad.
- Let it read your hardware. On first launch it checks RAM and chipset and narrows the model list to what will actually run well on your device, so you are not guessing.
- Download one model. Take the recommendation for your RAM tier rather than the biggest thing on the list. See which models run on your phone if you want to understand the choice first.
- Turn the network off and try it. Airplane mode is the honest test, and it is worth doing on day one so you know it genuinely works rather than trusting a claim.
How it compares to keeping ChatGPT
| ChatGPT and similar | On-device model | |
|---|---|---|
| Works with no connection | No | Yes, once the model is downloaded |
| Where your messages go | To the provider's servers | Nowhere; they stay on the handset |
| Account required | Yes | No |
| Cost | Free tier with limits, or a subscription | Free; the cost is storage and battery |
| Raw capability | Much higher | Lower, and honest about it |
| Speed | Network round-trip plus queue | Immediate, no round-trip |
| Knows today's news | Yes, with browsing | No, unless you enable optional web search |
| Battery impact | Minimal | Noticeable; it is real compute on your device |
Most people who do this end up using both: the cloud model for hard problems, the local one for everything private, offline or routine.
Questions people ask
Can I download ChatGPT to use offline?
No. OpenAI does not release the weights for its ChatGPT models, so there is nothing to download and run locally. What you can run offline are open-weights models such as Qwen, Gemma, Llama and Mistral, which are published for exactly this purpose.
Is there a free offline AI app?
Yes. TokForge is free, with no subscription and no account. Several other apps in this space are free too, and some are open source; the models themselves are free downloads from Hugging Face.
Does it really work in airplane mode?
Yes, once a model is on the device. That is the test worth running yourself on the first day.
Will it drain my battery?
Generating text is a sustained CPU or GPU load, so it uses meaningfully more power than reading or scrolling. We have not published a battery-drain measurement, so we will not put a number on it here. A long session will be noticeable.
How much storage does it need?
The app download is around 100 MB. Models are the large part, from roughly 0.4 GB for the smallest up to about 10 GB for a 14B model. You can delete and re-download models freely.
Does my data go anywhere?
No. Inference happens on the handset. TokForge has two features that can reach the network and both are off unless you turn them on: optional web search, and connecting to your own OpenAI-compatible server. Model downloads come from Hugging Face, and that is the only other time it uses the network.