Local AI: run AI models privately at home

You can run a capable AI chatbot on your own computer, for free, with no internet needed after setup. Your chats never leave your device. Here's what you need and how to start in about 10 minutes.

Mostly research-based. We did run Ollama with a small model on our own Linux server (CPU only, no graphics card); that output is shown below. We haven't tested the Mac, Windows or phone steps or the other apps yet. Facts come from official pages, last checked Oct 15, 2026.

Get LM Studio free →Get Ollama free →

Free plans and trials as of Oct 2026; some links may earn us a commission.

The short version

Easiest app: LM Studio (point-and-click) or Ollama (simple, popular).
Good first model: Qwen 3.5 4B or Gemma 4 E2B on an 8–16 GB computer.
The catch: local models are smaller than ChatGPT or Gemini, so expect simpler answers and slower replies on modest hardware.

Why local means privacy (and where it stops)

With a cloud chatbot, what you type is sent to the company's servers. With a local model, the AI runs on your processor and memory. After you download the model file, your prompts, answers and any documents you load stay on your device. LM Studio, for example, says it can “operate entirely offline” once you have model files.

Limits to know:

The download itself comes from the internet (the app's model library or Hugging Face).
Some apps have optional cloud features, usage stats or update checks; look in settings if you want them off.
Your device's own security still matters: anyone with access to your computer can read your chat history.
Local models don't search the web unless you add that, so their knowledge stops at their training date.

Not sure what your computer can handle? Use our free “What can my computer run?” checker →

What hardware do you need?

The rule of thumb: the model file must fit in memory with a few GB to spare. On a Mac, that's the unified memory (RAM). On a Windows or Linux PC with a graphics card, it's the card's video memory (VRAM); models can also run on the main processor and RAM, just more slowly.

Memory (Mac RAM or PC GPU VRAM)	Models that fit (download size, per Ollama)	What to expect
8 GB	Qwen 3.5 4B (3.3 GB), Gemma 4 E2B (4.6 GB), Llama 3.2 3B (2.0 GB)	Quick questions, summaries, simple writing
16 GB	Qwen 3.5 9B (6.6 GB), Gemma 4 12B (7.7 GB), DeepSeek-R1 8B (5.2 GB); gpt-oss 20B (14 GB) is a tight fit	Good everyday chat, light coding, reading documents
24–32 GB	Gemma 4 26B (16 GB), Qwen 3.5 27B (17 GB), Mistral Small 3.2 24B (15 GB), gpt-oss 20B	Noticeably smarter; closer to older cloud models
48–64 GB+	Qwen 3.6 35B (23 GB), DeepSeek-R1 70B (43 GB), gpt-oss 120B (65 GB, needs 80 GB+)	Enthusiast tier: big models, slower

Sizes are Ollama's default downloads (usually 4-bit compressed versions), checked Oct 15, 2026. “Fits” is our rule of thumb, not a guarantee; long documents need extra memory.

Mac: Apple Silicon (M1 or newer). LM Studio needs macOS 14+ and recommends 16 GB+; 8 GB Macs work with smaller models. Intel Macs aren't supported by LM Studio.
Windows: LM Studio recommends 16 GB RAM and at least 4 GB of dedicated VRAM, and needs a CPU with AVX2 (most from the last decade). It also supports Snapdragon X (ARM) laptops.
Linux: LM Studio ships as an AppImage (Ubuntu 20.04+); Ollama has a one-line installer.
Phones: PocketPal AI says you need 6 GB+ RAM for small models and 8 GB+ for better ones. Google's AI Edge Gallery needs Android 12+ or iOS 17+.

Beginner apps (as of Oct 2026)

App	Platforms	Why pick it	Cost
LM Studio	Mac (Apple Silicon), Windows, Linux	Friendliest point-and-click app; search and download models inside it. Version 0.4.26 per its changelog.	Free
Ollama	Mac, Windows, Linux	Hugely popular; one command to download and chat; many other apps plug into it.	Free for local use
Jan	Mac, Windows, Linux	Open-source ChatGPT-style app that runs offline.	Free, open source
GPT4All	Mac, Windows, Linux (x64)	Simple app from Nomic; can chat with your local documents. Minimum 16 GB RAM (8 GB for 3B models), per its GitHub.	Free, open source
PocketPal AI	Android, iPhone	Download small open models to your phone; works offline.	Free, open source
Google AI Edge Gallery	Android 12+, iOS 17+	Google's app for running Gemma models on your phone.	Free, open source

LM Studio website homepage with download buttons — LM Studio's homepage, where you download the app.Screenshot: lmstudio.ai, Oct 15, 2026 (public page, not from our testing)

Ollama model library page listing open models — Ollama's public model library, listing open models you can download.Screenshot: ollama.com, Oct 15, 2026 (public page, not from our testing)

Recommended open models (Oct 2026)

Qwen 3.5 / 3.6 (Alibaba): strong all-rounders from 0.8B to 122B; Qwen 3.6 focuses on coding and agent tasks, per Ollama's listing.
Gemma 4 (Google): E2B and E4B for laptops and phones, 12B for mid-range, 26B/31B for big machines.
gpt-oss (OpenAI): OpenAI's open-weight models: 20B (14 GB) and 120B (65 GB).
Llama 3.2 (Meta): tiny 1B/3B models that run almost anywhere. Meta's newest Muse models aren't confirmed as downloadable (not verified).
Mistral Small 3.2 (24B): good at following instructions, per Ollama's listing.
DeepSeek-R1: “reasoning” models that think step by step, 1.5B to 671B.

Licences differ: most of these allow personal use, but check each model's licence before using it in a business.

Platform guides: AI on Macs (Apple Intelligence, unified memory) · AI on Windows PCs (Copilot, Copilot+ PCs).

10-minute setup: Mac

Check you have an Apple Silicon Mac (Apple menu → About This Mac shows “M1” or later) and macOS 14 or newer.
Download LM Studio from lmstudio.ai (official site only) and drag it to Applications.
Open it, go to the model search (Discover), and pick a small model: Qwen 3.5 4B or Gemma 4 E2B on 8–16 GB.
Click download (a few GB; this is the only step that needs internet).
Open a new chat, choose the model, and ask: “Explain what you are in two sentences.”
Optional: turn Wi-Fi off and ask again. It should still answer.

Prefer Ollama? Download the Mac app from ollama.com, then open Terminal and type ollama run qwen3.5:4b.

10-minute setup: Windows

Check your memory: Settings → System → About (RAM), and Task Manager → Performance → GPU (dedicated memory).
Download LM Studio or Ollama for Windows from the official site and run the installer.
LM Studio: search for Qwen 3.5 4B (or Gemma 4 12B if you have a 12 GB+ graphics card), download, and start a chat.
Ollama: open PowerShell and type ollama run qwen3.5:4b. The first run downloads the model, then you can chat.
Type /bye to exit Ollama's chat.

10-minute setup: Linux or Intel Mac

Install Ollama: on Linux run the one-line installer from ollama.com/download; on an Intel Mac download the Mac app there. (LM Studio also offers a Linux AppImage, but doesn't support Intel Macs.)
Open a terminal and type ollama run qwen3.5:4b. The first run downloads the model.
Type a question and press Enter. On older or CPU-only machines, try a smaller model like qwen3.5:0.8b.

What we actually ran (Oct 15, 2026)

On our Linux server (8 CPU cores, no graphics card), we installed Ollama 0.40.4 and ran Qwen 3.5 0.8B, the smallest Qwen 3.5. Our server's network blocked Ollama's own download servers, so we downloaded Qwen 3.5 0.8B as a GGUF file from Hugging Face (ggml-org) and imported it into Ollama. It answered in about 1.5 seconds at about 29 tokens (word pieces) per second. The answer was sensible but simple, which is typical for a model this small.

Terminal showing ollama version 0.40.4, the model list, a prompt about privacy, the model's answer and timing statistics — Real terminal output from our test: Ollama running Qwen 3.5 0.8B and printing speed stats.Our test: Ollama 0.40.4 on our Linux server, CPU only, Oct 15, 2026

Local vs cloud: honest tradeoffs

	Local	Cloud (ChatGPT, Gemini, Claude)
Privacy	Stays on your device	Sent to the company's servers
Quality	Good for everyday tasks; weaker on hard reasoning	Best available models
Speed	Depends on your hardware	Usually fast
Cost	Free (plus electricity and your hardware)	Free tiers; $8–$20+/month for more
Web search, agents	Not built in	Built in

Safety tips

Download apps only from official sites (ollama.com, lmstudio.ai, jan.ai, nomic.ai) or the official app stores. Fake “AI app” downloads are a common malware trick.
Get models from the app's built-in library or well-known Hugging Face publishers.
A local model can still be wrong; double-check anything important.
Keep the app updated, and don't expose its local server to the internet unless you know what you're doing.

FAQ

Is it really free? The apps above are free for personal local use, and open models are free to download.

Will it slow my computer down? While it's answering, yes, it uses a lot of memory and processor. Close it when you're done.

Can I use it for AI agents? Partly. Some agent tools can connect to Ollama or LM Studio, but local models are less reliable at multi-step tasks. See our agentic AI hub.

Which model is “best”? The biggest one that fits your memory comfortably. Start small and move up.

Does it work on an old Intel Mac? LM Studio says Intel Macs aren't supported; try a cloud chatbot instead.

More: AI models explained · Best free AI models · Agentic AI