To run Qwen 3.5 on a phone, install an app that loads open models, download one of the phone-sized versions once, and chat with no connection. Qwen 3.5 4B is the best starting point: a 2.74 GB download that runs on phones with 3 GB of RAM or more. Phones with 6 to 8 GB can step up to Qwen 3.5 9B, which gives noticeably better answers. Everything runs on the phone, so nothing you type goes to Alibaba or anyone else.
What is Qwen 3.5? #
Qwen 3.5 is the current generation of open-weight models from Alibaba’s Qwen team. The family runs from sub-1B models meant for phones up to very large models built for servers, and the small ones are what matter here.
What sets it apart on a phone, per the model cards:
- Very wide language coverage, over 200 languages and dialects, which makes it one of the most multilingual families you can run locally.
- Vision built in. Qwen 3.5 is trained on text and images together rather than having vision bolted on.
- Thinking mode. The models can write out step-by-step reasoning before answering.
- A long native context window, though a phone uses far less of it than a server would.
- Apache 2.0 license, so it’s free for personal and commercial use.
Which Qwen 3.5 size fits your phone? #
| Model | Download | Needs RAM | What it’s like |
|---|---|---|---|
| Qwen 3.5 0.8B | 0.81 GB | 2 GB+ | Very fast, limited quality. For low-end phones |
| Qwen 3.5 4B | 2.74 GB | 3 GB+ | The sweet spot for everyday phones |
| Qwen 3.5 9B | 5.68 GB | 6 GB+ | The best answers a phone can give, slower |
Sizes are from the Personal LLM catalog, which carries the 0.8B, 4B and 9B. The app reads your phone’s RAM and shows “Fits your device” when a model has at least 2 GB of headroom. Other sizes in the family, or different quantizations of these, can be added as your own GGUF by URL. If you’re unsure what your phone has, how much RAM you need to run an LLM lists the common tiers.
How to run Qwen 3.5 on iPhone or Android #
- Install Personal LLM from the App Store or Google Play. It’s free, needs no account, and runs on iOS 15.1 or later and Android 7 or later.
- Open the Models tab. Qwen 3.5 4B is marked as the recommended first download.
- Download it on Wi-Fi. The file comes directly from Hugging Face. Interrupted downloads resume, and the app checks storage before starting.
- Open a new chat and pick Qwen 3.5 in the chat header.
- Switch on airplane mode and ask something. It keeps working, because nothing needs the network.
- Add image support (a separate 195 MB to 1 GB download) if you want to ask about photos.
To try a different size or quantization, add it as your own GGUF by URL. Run a custom GGUF model on your phone walks through finding the right file and link.
Should you turn on thinking mode? #
Qwen 3.5 reasons by default under its standard settings, writing its working inside think tags before the answer. In Personal LLM, thinking mode is off by default, because most questions don’t need it and replies come faster without it. Turn it on per chat, or tap the Thinking preset, for:
- Math and word problems
- Logic puzzles
- Code and debugging
- Anything where you want to check the model’s working
The reasoning appears in a collapsible panel above the answer, so it doesn’t clutter the chat. It costs time and tokens, so leave it off for rewriting, summaries and quick questions. Can local AI do math shows the difference it makes, and AI thinking mode explained covers when the reasoning is worth trusting.
Using Qwen 3.5 in other languages #
This is where Qwen stands out. With support for over 200 languages and dialects, it handles Chinese, Spanish, Arabic, Hindi, Indonesian, Vietnamese and many less common languages better than most models of its size. You can:
- Chat entirely in your own language
- Translate a message and ask for a more formal or casual version
- Paste text in one language and ask for a summary in another
- Practice a language you’re learning, with corrections
Two tips. Write your question in the language you want the answer in, or say “Answer in Portuguese.” And expect the 9B to be clearly better than the 4B in less common languages, where smaller models thin out fast. For more, see offline AI in other languages.
Asking Qwen 3.5 about photos #
With image support added, attach a photo or take one, and ask. Qwen 3.5 can describe a scene, read text in an image, explain a chart, or tell you what a form is asking for. Photos stay on the phone. On phones with 8 GB or more, GLM 4.6V Flash is another strong option for documents and screenshots.
Small vision models misread dense or stylized text, so check anything that matters.
How much context can Qwen 3.5 use on a phone? #
Far less than its native window, because every token of context costs memory and slows replies. Personal LLM starts at 2,048 tokens, caps chats with images at 4,096, and lets you raise the setting if your phone has spare RAM. For long documents, attach them as files instead: the app searches the document on the phone and gives the model only the relevant passages. Why does AI forget what I said explains the trade-off.
Is Qwen safe to run on your phone? #
Running it locally removes the main privacy question people have about Chinese AI services, because there’s no server involved: the model is a file on your phone and your chats never leave it. The model’s training can still shape its answers on some topics, which is worth knowing rather than worrying about. Both sides are covered in is Qwen safe to use.
Frequently asked questions #
Can Qwen 3.5 run on an iPhone? #
Yes. Qwen 3.5 4B runs on iPhones with about 4 GB of RAM or more, and the 9B on iPhones with 8 GB, such as the iPhone 15 Pro and later. The app uses the iPhone’s GPU through Metal.
Does Qwen 3.5 work without internet? #
Yes. After the one-time download it runs completely on your phone. It won’t know about recent events, since it has no web access, but everything else works offline.
Which is better on a phone, Qwen 3.5 or Gemma 4? #
Qwen 3.5 9B gives the strongest answers if your phone can run it, and the 4B is a good all-rounder; Gemma 4 E2B is lighter and quicker on modest phones. Our head-to-head, Qwen 3.5 vs Gemma 4 on your phone, compares them properly, and you can install both and switch mid-chat.
Is Qwen 3.5 free to use? #
Yes. The weights are released under the Apache 2.0 license and cost nothing to download, and Personal LLM is free with ads in the free version.
Why is Qwen 3.5 0.8B giving bad answers? #
It’s a very small model meant for phones with little memory. It’s fast, but it knows much less and makes more mistakes. If your phone has 3 GB of RAM or more, the 4B is a large step up.