How to run free AI on your phone

Check your device memory, pick a model that fits, and have it running in a few minutes. After the download, it works with no account and no internet.

How much RAM you need

RAM (memory) decides which models your phone can load. The model file has to share memory with the operating system and the app, so we do not treat the full advertised RAM as usable.

On iOS we treat about 57% of device RAM as usable (in the 55–60% range), because the system keeps a large share and may evict an app that goes higher.

On Android we treat about 67% of device RAM as usable (in the 65–70% range), leaving room for the system and the app.

A model is marked as a fit when its Q4_K_M file, or a smaller Q3_K_M file, stays within that budget. These are rules of thumb, not a guarantee. Close other apps before you load a model.

How to check RAM on iOS: open Settings → General → About, or look up the RAM for your device class. On Android: open Settings → About phone and look for RAM or Memory (on Samsung: Settings → About phone → Phone information, or Device care).

On the directory, pick iOS or Android and a RAM amount in the “Can my phone run this?” bar. If this browser reports a memory hint (Chrome on Android often does; iOS Safari usually does not), we may pre-select a tier and label it as a guess. That hint is often capped at 8 GB.

RAMModels that fit (size rule of thumb)
4 GB24 models →
6 GB56 models →
8 GB66 models →
12 GB127 models →
16 GB+135 models →

Run it on your phone

Install an app that can load a GGUF file, then download the file from the model page.

iOS: PocketPal AI, Layla.

Android: PocketPal AI, Layla, ChatterUI.

PocketPal AI is free on both stores and can search Hugging Face by repository ID. ChatterUI is Android-only and is installed from its GitHub releases. Layla is paid on iOS.

  1. Install one of those apps for your operating system.
  2. Open a model page here and copy the Hugging Face repository ID and the Q4_K_M file name.
  3. In the app, add a model from Hugging Face, paste the repository ID, and download the Q4_K_M file, or Q3_K_M if memory is tight. Use wifi.
  4. Load the file and chat. From now on it works fully offline.

Scores and speed estimates

These scores come from the public UGI (Uncensored General Intelligence) Leaderboard on Hugging Face, and were first listed here from the uncensored-ai-phone directory. Uncensored IQ is described there as a blend of knowledge and how freely a model shares it, Reasoning as problem-solving, and Answers (out of 10) as how willing the model is to respond. We do not have a separate methodology and we do not re-run the tests.

The directory also shows an estimated tokens per second range by model size, labelled as an estimate. The ranges are a rough rule for a Q4_K_M file on a recent flagship chip (Snapdragon 8 Gen 3 or A17 Pro class): about 25–45 tok/s under 1.2B parameters, 18–30 tok/s under 2.2B, 12–22 tok/s under 5B, 8–16 tok/s under 8B, and 6–12 tok/s above that. They are not measurements of the listed models.

Quantization in one minute

Models are shrunk for phones by storing each weight in fewer bits. Q4_K_M (about 4.8 bits per weight) is the standard choice. Q3_K_M is roughly 20% smaller with a small quality drop. IQ files are newer, more compact formats; Q5–Q8 and F16 are larger and mainly for computers.

Frequently asked questions

Does it cost anything?

No. The model files and PocketPal are free. There is no subscription, no API fee and no account. Some other apps, such as Layla, are paid. Licences differ between models (Apache 2.0, Gemma, Llama and a few non-commercial ones), so check the licence on a model's page before using it in a product.

Can a phone really run an AI model with no internet?

Yes. Every model here is a small language model that has been compressed (quantized) into a single GGUF file of roughly 0.5 to 6 GB. You download it once, and from then on your phone's own processor generates every answer. It keeps working in airplane mode, and nothing you type is sent anywhere.

Why run a model on your phone instead of a cloud chatbot?

It is free after the download, and it stays private. Your notes, ideas and questions never leave the device. There is no account, no subscription and no rate limit, and it still works on a plane, in the subway or anywhere without a signal.

How do I know if a model fits my phone?

Pick iOS or Android and your device memory in the bar at the top of the list. We treat only part of the RAM as usable, because the operating system and the app need the rest: about 57% on iOS (in the 55–60% range) and about 67% on Android (in the 65–70% range). A model counts as a fit when its Q4_K_M file, or a smaller Q3_K_M file, stays within that budget. These are rules of thumb, not a guarantee. Close other apps before you load a model, and if it will not load, try the smaller file or a smaller model.

How capable are phone-sized models?

Be realistic: they are noticeably weaker than large cloud or desktop models. The strongest uncensored model in this list scores 44.7 on the UGI Leaderboard, and the tiniest ones are best kept to quick lookups. They are useful for brainstorming, summarising and talking an idea through, but double-check anything that matters.

What is the difference between Official and Uncensored?

Official models are mainstream releases from the original authors, or GGUF conversions by known quantizers such as bartowski, unsloth, lmstudio-community and ggml-org. Uncensored models are community fine-tunes that reduce refusal behaviour. Official is the default view. The uncensored list is one tap away.

What does "uncensored" mean when it appears in a model name?

Some authors have stripped out or weakened the refusal behaviour that ships with the original model, so it tries to answer instead of replying "I can't help with that". Common techniques are abliteration, the Heretic tool and targeted fine-tuning, which is why you see words like abliterated, heretic or uncensored in the names. You are responsible for how you use the output.

What do Uncensored IQ, Reasoning and Answers mean?

They are published scores from the public UGI (Uncensored General Intelligence) Leaderboard on Hugging Face, and were first listed for this directory from the uncensored-ai-phone project. Uncensored IQ is described there as a blend of how much a model knows and how freely it shares it. Reasoning is described as how well it works through problems. Answers, out of 10, is described as how willing it is to respond at all. We do not have a separate, independently documented methodology, and we do not re-run the tests. Official models and some newer uncensored models have no UGI scores, so we show downloads and likes instead.

How should I read the tokens per second figures?

They are estimates, not measurements from this site. The ranges are a rough rule by model size on a recent flagship chip (roughly Snapdragon 8 Gen 3 or A17 Pro class, running a Q4_K_M file). A 1B-class model is often around 25–45 tokens/s, 3–4B around 12–22, and 8B-class around 6–14. Older chips, thermal throttling, long context and other apps will be slower. Treat the figure as a planning aid only.

What is GGUF, and what do Q4_K_M and Q3_K_M mean?

GGUF is the file format used by llama.cpp, the engine inside apps like PocketPal and ChatterUI. The Q-number is how many bits each weight is squeezed into: Q4_K_M is a 4-bit medium mix and the usual sweet spot of size and quality, Q3_K_M is about 20% smaller and slightly less accurate, while Q5, Q6 and Q8 files are bigger and closer to the original.

The model is slow or will not load. What can I do?

Close other apps to free memory, update the app, or switch to the Q3_K_M file or a smaller model. The first reply after loading is always the slowest, and phones throttle when they get hot. Newer chips are dramatically faster.

Is this safe? Can a model act or trade on its own?

The files are hosted on Hugging Face, the main public home for open models, and downloaded through apps such as PocketPal. A model on its own only produces text: it cannot browse, see live prices, place trades or touch your accounts. It can also be confidently wrong, so treat its output as a starting point, not advice.

Copied