Your personal AI, on your PC
Choose a trusted free tool, match a model to your hardware, and follow one clear step at a time. Your first private conversation is closer than it looks.
You need internet to install the app and model files, which can be several gigabytes.
After setup, ordinary chats can run offline. The model uses your computer’s memory and processor.
Your prompts stay local unless you choose a cloud feature, synced folder, or outside service.
Local AI is valuable because you control it—not because it is identical to a large paid service.
A local assistant can be excellent for drafting, summarizing, brainstorming, coding help, and private document work. But it will usually be slower and less capable than Meta Muse or leading paid cloud models, especially on an everyday laptop.
Work through the five steps. Your progress stays only until this page is closed or refreshed.
You can switch later without losing your model knowledge.
This browser can estimate your operating system, CPU threads, and—on some browsers—memory.
Download GPT4All from Nomic AI’s official website, then open the installer.
Use the official download page. Open the downloaded installer and follow its prompts.
A model is the part that writes replies. Start smaller than you think—you can add another later.
Fast and lightweightFor computers with 8 GB RAM. Quicker, with lower model quality.7B-class modelFor 16 GB RAM. Choose an in-app “Best overall” option.7B-class or largerFor 32 GB+ and stronger hardware. Start with 7B before going bigger.Open GPT4All’s in-app model library. Choose Fast and lightweight for an 8 GB computer, or a Best overall 7B-class model for 16 GB or more. Select Download once, then start chatting.
GPT4All: Use the in-app descriptions. Choose a “Fast and lightweight” model with 8 GB RAM, or a “Best overall” 7B-class model with 16 GB or more.
Ollama or LM Studio · Light: llama3.2:1b, qwen3:0.6b, qwen3:1.7b, gemma3:1b
Ollama or LM Studio · Standard: qwen3:8b, llama3.1:8b, gemma3:12b, phi4
Ollama or LM Studio · Power: qwen3:32b, deepseek-r1:32b, llama3.3:70b
Rule of thumb: a quantized 8B model needs roughly 5–6 GB of available memory. A 70B model generally needs 40 GB or more. For slower, explicit reasoning try deepseek-r1:8b. For coding, Qwen 3 is a good first choice; deepseek-coder-v2 and codellama are alternatives.
GPT4All is a standalone app, so it does not need a local server or connection test.
Your first prompt
“You are my private writing partner. Help me improve a short message while preserving my meaning and tone. Ask one question before rewriting it.”
Copy a proven starting instruction, then replace the bracketed text with your own material.
Talk through decisions without sending the conversation to a cloud model.
Draft, revise, shorten, or adjust tone. Always review names, dates, numbers, and claims.
Generate options, then make the model challenge its own first ideas.
Turn notes you already collected into themes, questions, and a source-aware outline.
Ask for one self-contained HTML file. Save the reply as index.html, then double-click it to open in your browser.
Start with a browser app that needs no installation. Work in a new folder, keep backups, and test each change before asking for the next one.
index.html and double-click it for a local preview.Limit: Small local models often lose track of large codebases and can create insecure or broken code. Use them for small prototypes—not security-sensitive, financial, medical, or production systems without expert review.
GPT4All’s built-in LocalDocs is the simplest option: point it at a folder and ask questions without uploading anything.
GPT4All · easiest: Open LocalDocs, choose a folder containing PDFs, Word documents, or text files, and let GPT4All index it locally. Then select that collection in Chat and ask questions in natural language. The files and index stay on your computer.
Ollama + Open WebUI · flexible alternative:
pip install open-webuiopen-webui serveollama pull nomic-embed-textOpen http://localhost:8080, connect Ollama, select nomic-embed-text for embeddings, then upload documents to a workspace. AnythingLLM is another open-source desktop alternative focused on document Q&A.
Updates are easier when you change one thing at a time.
Use the app’s built-in updater or download the latest version from the same official site. Close the app before installing an update.
Keep the model that works. GPT4All users can return to the in-app model library. For Ollama, use ollama pull <model> to download or refresh and ollama list to see installed models.
Chat history locations vary by tool. Check the official documentation before uninstalling, moving folders, or clearing application data.
Start with the simplest explanation. Most setup problems do not require technical expertise.
Privacy and accuracy are separate. Local processing helps with the first; careful judgment is still needed for the second.
Models can invent confident-sounding details. Check medical, legal, financial, safety, and current information with a qualified or primary source.
Use official product sites and model names from the official library. Avoid repackaged installers and unfamiliar mirrors.
Local chats stay on the computer, but operating-system backups or synced folders may copy application data to the cloud.
Local models may produce inaccurate or unsuitable content. Use them with active adult supervision rather than as an authority.
The few technical words you are likely to meet, translated into plain English.
nomic-embed-text does this locally.