# How to run an LLM locally on Mac, Windows, iPhone and Android > How to run an LLM locally on Mac, Windows, iPhone or Android: what you need, step-by-step setup, the best apps, real speeds, and honest limits of local AI. Canonical: https://askloci.ai/guides/run-llm-locally/ Last updated: 2026-11-03 Running a large language model (LLM) locally means the model runs on your own device instead of on a company's servers. After a one-time download, you can chat with it with no internet connection, no per-message fee, and no need to send your text to a cloud service. A few years ago that required a developer's setup. Today it takes an app and a download. This guide explains what you need, the main ways to do it on each device, and what to expect. We make Loci, one of the apps listed, and we say so where it comes up. ## What you need 1. **A device with enough memory.** This is the key limit. Models are listed by download size, and the model has to fit in memory alongside your system. For reference, Ollama's library lists Gemma 4 E2B at 4.6-7.5 GB, Qwen3.5 4B at 3.3-4.0 GB and gpt-oss 20B at 14 GB. 2. **An app that runs models.** See the table below. 3. **A model.** Open-weight families include Gemma 4 (Google), Qwen3.5 (Alibaba) and gpt-oss (OpenAI). Our guide to the [best local LLMs](https://askloci.ai/guides/best-local-llms/) lists sizes and hardware tiers. 4. **Free storage and a one-time download.** Models are multi-gigabyte files. Download on Wi-Fi. ## The main options at a glance | App | Mac | Windows | iPhone/iPad | Android | Linux | Picks a model for you | Cost | | --- | --- | --- | --- | --- | --- | --- | --- | | Loci | Apple silicon, macOS 26.2 or later (beta) | Windows 11, x64 and ARM64 (beta) | Yes | Yes | No | Yes | Free | | LM Studio | Apple silicon | Yes | Companion app (Locally) | No | Yes | No | Free | | Ollama | Yes | Yes | No | No | Yes | No | Free locally | | Jan | Apple silicon | Yes | No | No | Yes | Yes | Free | | GPT4All | Yes | Yes | No | No | Ubuntu | No | Free | | PocketPal | No | No | Yes | Yes | No | No | Free | Sources: each app's own documentation and our comparison page; all rows on the publish date. ## On a Mac **What works:** any Mac with Apple silicon (M1 or newer). Intel Macs are poorly supported by current apps: Jan, for example, no longer runs local models on Intel Macs. **Easiest route (Loci or Jan):** 1. Download the app for macOS. Loci's direct Mac download requires Apple silicon and macOS 26.2 or later. 2. Open it. It downloads a model that fits your Mac's memory. 3. Ask a question. Turn off Wi-Fi to confirm it still works. **More control (LM Studio):** install, open **Discover**, search for a model, pick a size that fits, download, load, chat. **Developer route (Ollama):** download the app from ollama.com/download or install from the terminal, then run for example `ollama run gemma4:e4b`. **What to expect:** in Loci's published data, an M2 Mac with 8 GB ran Qwen3.5 2B at a median of 44 tok/s, an M5 with 16 GB ran Gemma 4 E4B at 33 tok/s, and an M5 Pro with 24 GB ran Gemma 4 12B at 31 tok/s. These are from 2 to 4 Macs each, so treat them as indications. See also our [Mac page](https://askloci.ai/local-llm-mac/). ## On Windows **What works:** Windows 11 PCs. Apps run models on the CPU, GPU or NPU depending on the app and hardware. **Easiest route:** Loci for Windows (x64 and ARM64; labelled beta) and Jan both set up a model for you. **More control:** LM Studio and GPT4All both have Windows installers. GPT4All offers a Windows ARM version. **Developer route (Ollama):** download the installer from ollama.com/download or run `irm https://ollama.com/install.ps1 | iex` in PowerShell, then `ollama run gemma4:e4b`. **Tip:** if you'd just like to confirm the app is fully offline, install it, let it download a model, then disconnect from Wi-Fi. **What to expect:** we do not yet have a published Windows speed table. Don't trust a number that claims to cover "Windows PCs" without naming the chip and memory. ## On an iPhone or iPad **What works:** recent iPhones. Current medians, sample sizes and answer counts are on the [speed pages](https://askloci.ai/speed/). **Options:** - **Loci:** free, no account; picks a model for your phone. Loci Link lets your iPhone use bigger models running on your Mac or PC, without an account. - **Locally (LM Studio's app):** connects to models on your own computer, needs an LM Studio account for LM Link. - **PocketPal:** an open-source phone app for small models on-device. - **Private LLM:** a paid on-device app for iPhone and iPad. **Steps (Loci):** install from the App Store; open; let the model download over Wi-Fi; ask a question; try airplane mode. ## On Android **What works:** recent mid-range and flagship phones. Speeds vary widely by phone. Current medians are on the [speed pages](https://askloci.ai/speed/). Under about 5 tok/s feels slow to read. **Options:** Loci (free), PocketPal, and AnythingLLM's Android app, which pairs with its desktop app. **Steps (Loci):** install from Google Play; open; wait for the model download on Wi-Fi; ask a question. ## What it's good at, and what it isn't **Good at:** drafting and rewriting, summarising, brainstorming, explaining things, simple coding help, and working with files you give it, all private and offline. **Weaker at:** hard multi-step reasoning, very long documents, obscure facts and anything that changes by the day. Local models are far smaller than the biggest cloud models. If you need live information, switch on web search where an app provides it; that request does use the internet. In Loci, it's one switch, and requests go directly to search providers. Vendor benchmarks show big differences even within one family: Google reports 85.2% on MMLU Pro for Gemma 4 31B and 69.4% for the E4B. Those figures are in our [best local LLMs](https://askloci.ai/guides/best-local-llms/) guide. ## Troubleshooting - **The app is slow or the phone gets warm.** Pick a smaller model; close other apps. - **The download fails.** Check free storage and use Wi-Fi. - **The answers are poor.** Try a bigger model if your device can run it; ask more specific questions; or use web search for current facts. - **It says the model won't fit.** That's the app protecting you. Choose a smaller size. ## FAQ ### What does it mean to run an LLM locally? The model's weights are stored and run on your own device, so prompts and answers don't need to go to a server. ### Can I run an LLM locally for free? Yes. LM Studio, Jan, GPT4All, Ollama's local runtime, PocketPal and Loci are free for local use. Some of them offer paid cloud features. ### What hardware do I need to run an LLM locally? Mostly memory. About 8 GB can run small 2B-4B models; 16 GB runs 9B-12B; 24 GB or more runs 20B-31B. These are our rough estimates based on the sizes Ollama lists. ### Can I run a local LLM on my phone? Yes. iPhone, iPad and Android apps like Loci and PocketPal run small models on-device. ### Is there a private AI app for iPhone that works offline? Yes. Loci runs the model on the iPhone itself. After the download, typed chat works with no connection and no account. ### Do I need an internet connection? Only to download the model, to use web search and to download updates. Typed chat works offline after the download. ### Is a local LLM private? Local chats stay on your device. Anything online (web search, cloud features) can leave it, so check settings. ### Which is the best local LLM app? It depends on the device and how much control you want. See [Ollama alternatives](https://askloci.ai/compare/ollama-alternative/) and [LM Studio alternatives](https://askloci.ai/compare/lm-studio-alternative/).