Run AI on your own computer
Private, offline, and free once you own the machine. Pick your computer to see the best AI it can run for each job, and how hard the setup is.
What can my computer run?
Not sure? On Windows, open Task Manager, then Performance: it shows your memory and your graphics card. On a Mac, Apple menu, then About This Mac.
9 of 17 kinds of AI run on any laptop, 8 GB RAM, no graphics card. Chat models: 1B to 4B; about 7 to 9 tokens a second, near reading speed. Online only on this machine: image generators, upscalers, 3d generators, video generators, talking avatars, music generators, coding agents, app builders.
9 of 17 kinds of AI run on laptop or desktop, 16 to 32 GB RAM, no graphics card. Chat models: 7B to 9B; about 5 to 13 tokens a second; 30B MoE models usable with 32 GB. Online only on this machine: image generators, upscalers, 3d generators, video generators, talking avatars, music generators, coding agents, app builders.
11 of 17 kinds of AI run on apple Silicon Mac, 16 GB. Chat models: 8B to 12B; about 10 to 20 tokens a second, faster than you read. Online only on this machine: 3d generators, video generators, talking avatars, music generators, coding agents, app builders.
13 of 17 kinds of AI run on apple Silicon Mac, 24 to 36 GB. Chat models: Up to 27B to 30B; dense 27B about 5 to 12 tokens a second, 30B MoE much faster. Online only on this machine: 3d generators, video generators, talking avatars, music generators.
13 of 17 kinds of AI run on apple Silicon Mac, 64 GB or more. Chat models: Up to 70B; about 6 to 8 tokens a second, gpt-oss-120b with 96 GB or more. Online only on this machine: 3d generators, video generators, talking avatars, music generators.
13 of 17 kinds of AI run on gaming PC or laptop, 8 GB graphics card. Chat models: 7B to 8B; about 50 tokens a second, very fast. Online only on this machine: 3d generators, video generators, coding agents, app builders.
14 of 17 kinds of AI run on gaming PC, 12 GB graphics card. Chat models: 12B to 14B; about 30 to 55 tokens a second. Online only on this machine: video generators, coding agents, app builders.
15 of 17 kinds of AI run on gaming PC, 16 GB graphics card. Chat models: 14B dense, about 40 to 80 tokens a second; gpt-oss-20b fits fully. Online only on this machine: coding agents, app builders.
17 of 17 kinds of AI run on high-end gaming PC, 24 to 32 GB graphics card. Chat models: 27B to 32B; about 35 to 60 tokens a second.
17 of 17 kinds of AI run on workstation with several graphics cards. Chat models: 70B dense about 15 tokens a second; 120B MoE like gpt-oss-120b with 80 GB.
The apps that make it easy.
You need one app to run the models. Start with the first one.
Easiest app to download and chat with local models
Free for home and work since July 2025.
Get LM StudioOpen source ChatGPT-style app for local and online models
Free and open source.
Get JanPolished chat app mixing local and online models
Free plan; Aurum $149 a year or $349 lifetime.
Get Msty StudioChat with your own documents using local models
Desktop app free and open source.
Get AnythingLLMSimple local chat app, no longer updated often
Free and open source; last release February 2025.
Get GPT4AllRuns small chat models offline on your phone
Free and open source.
Get PocketPal AILocal image and short video generation on Apple devices
Free edition runs locally; Draw Things+ $8.99/mo for cloud compute.
Get Draw ThingsOne-click installer and manager for image generation apps
Free and open source.
Get Stability MatrixSimple Mac image generator, not updated since 2024
Free; last release August 2024, older models only.
Get DiffusionBeeOne-click installer for hundreds of open AI apps
Free and open source.
Get PinokioMake photos sharper and bigger, offline
Free and open source; last desktop release December 2024.
Get UpscaylTranscribe audio and meetings on a Mac, offline
Free version; Pro €64 once, lifetime updates.
Get MacWhisperFree Whisper transcription app for any computer
Free and open source.
Get BuzzFree offline transcription app with subtitles export
Free and open source.
Get VibeFree offline dictation, speak and it types
Free and open source.
Get HandyRuns local models in the background for other apps
Local use free and unlimited; optional cloud plans from $20/mo.
Get OllamaOne-file app for local chat and story writing
Free and open source.
Get KoboldCppPolished canvas app for local image generation
Community Edition free and open source.
Get InvokeChatGPT-like web page on top of Ollama
Free; its license requires keeping the Open WebUI branding.
Get Open WebUIThe engine inside most local chat apps, command line
Free and open source (MIT).
Get llama.cppNode editor for local image, video and audio models
Free and open source (GPL-3.0); Comfy Cloud is paid.
Get ComfyUISelf-hosted server for chat, image and voice models
Free and open source.
Get LocalAIHow much computer do I need?
For chat models. Images need a little less, video a lot more (usually an NVIDIA card).
| Computer | Chat model that runs well |
|---|---|
| Any laptop, 8 GB RAM, no graphics card | 1B to 4B; about 7 to 9 tokens a second, near reading speed |
| Laptop or desktop, 16 to 32 GB RAM, no graphics card | 7B to 9B; about 5 to 13 tokens a second; 30B MoE models usable with 32 GB |
| Apple Silicon Mac, 16 GB | 8B to 12B; about 10 to 20 tokens a second, faster than you read |
| Apple Silicon Mac, 24 to 36 GB | Up to 27B to 30B; dense 27B about 5 to 12 tokens a second, 30B MoE much faster |
| Apple Silicon Mac, 64 GB or more | Up to 70B; about 6 to 8 tokens a second, gpt-oss-120b with 96 GB or more |
| Gaming PC or laptop, 8 GB graphics card | 7B to 8B; about 50 tokens a second, very fast |
| Gaming PC, 12 GB graphics card | 12B to 14B; about 30 to 55 tokens a second |
| Gaming PC, 16 GB graphics card | 14B dense, about 40 to 80 tokens a second; gpt-oss-20b fits fully |
| High-end gaming PC, 24 to 32 GB graphics card | 27B to 32B; about 35 to 60 tokens a second |
| Workstation with several graphics cards | 70B dense about 15 tokens a second; 120B MoE like gpt-oss-120b with 80 GB |
Speeds from public llama.cpp benchmarks. A token is about three quarters of a word; people read 5 to 8 tokens a second.
Quick answers.
Can I run ChatGPT on my own computer?
Not ChatGPT itself, but open models that are close to it. Gemma runs on a 16 GB laptop with a free app like LM Studio. No account, works offline, nothing leaves your machine.
How much RAM do I need to run an LLM locally?
8 GB runs small models (1B to 4B). 16 GB runs good everyday models (8B to 12B). 32 GB or a 24 GB graphics card runs strong 27B to 32B models. 64 GB and up runs 70B.
Is running AI locally free?
Yes. The apps and most open models are free to download and use, including for work in many cases. You pay only for the computer and electricity. Check each model’s license on its page.
Mac or PC for local AI?
A Mac with Apple chip and lots of memory is the simplest and quietest. A PC with an NVIDIA graphics card is faster for images and video, and many video models need NVIDIA.
What is the easiest way to start?
Install LM Studio, search for a model it recommends for your computer, click download, and start chatting. It takes about ten minutes.