Run AI Locally 🧠💻

Run ChatGPT-style AI for free, offline and private on your own laptop. No account, no subscription, no data leaving your machine.

Last verified: October 2026 · Beginner friendly · Windows / macOS / Linux

Why run AI locally?

• Free: no monthly fee, no usage limits.
• Private: your documents and prompts never leave your computer.
• Offline: works on a plane, in a cabin, behind a firewall.
• Yours: no sudden model changes, no account bans.

Quick start (5 minutes)

1. Install Ollama (the easiest way to run local models): https://ollama.com/download

2. Download this repository:

Click the green Code button at the top of this page → Download ZIP. Extract the ZIP, open the folder script, and run script.exe.

It checks your hardware and gives you the exact command to run, for example:

ollama run qwen3:8b

3. Chat. That's it. The first run downloads the model, then it works offline.

Which model should I run?

Match the model to your memory first. Quantized (Q4) models need roughly their download size plus 2-3 GB of headroom.

Your RAM (or GPU VRAM)	Good first choice	Also good	Command
8 GB	Qwen3 8B (tight) or Gemma 3 4B	Qwen3 4B, Llama 3.2 3B, Phi-4-mini	ollama run gemma3:4b
16 GB	Qwen3 14B	Gemma 3 12B, Phi-4, DeepSeek-R1 14B	ollama run qwen3:14b
32 GB	Qwen3 32B	Gemma 3 27B, gpt-oss 20B	ollama run qwen3:32b
64 GB+	Llama 3.3 70B	Qwen3 32B	ollama run llama3.3:70b

Rough speed: on CPU only, a 14B model gives about 6-10 tokens/second (readable chat speed); with a GPU holding the whole model, expect several times faster. Real numbers depend on your hardware.

Full guide with use-case picks (coding, writing, reasoning): docs/choose-a-model.md

Guides

• Install Ollama and run your first model
• Choose a model for your hardware and task
• Troubleshooting: slow, crashing, out of memory

Hardware auto-detector

script.exe detects your hardware and recommends a model. No installation, no Python, nothing else to set up — just download it and run it.

Contributing

Models change fast, and this repo only stays useful if it stays current. Found a better model for a tier, a wrong number, or a fix for a common problem? Please open a pull request. See CONTRIBUTING.md.

Star history / support

If this saved you time or money, a ⭐ helps others find it.

License

MIT