Your own AI, on your own terms

What running your own AI gets you

• Choice: pick the model, the version and its standing instructions. Nobody swaps it out from under you.
• Privacy, if you set it up right: a model on your own machine doesn't need to send your words anywhere.
• Independence: no account, no price change, no plan cut, works offline.
• Control of behavior: the maker's model, or a community retune that follows your instructions instead.

And what it costs: weaker answers on hard jobs unless you have serious hardware, setup and upkeep, and security that is now your job. The full list, both columns, is in Local or hosted: the honest trade-offs.

Good first models

• Fits 8GB

  Gemma 4 E4B

  The best first model for an 8GB machine. Apache 2.0.

  ollama run gemma4:e4b

• Fits 16GB

  Gemma 4 12B

  The sweet spot for a 16GB laptop: documents, images, writing. Apache 2.0.

  ollama run gemma4:12b

• Fits 24–32GB

  Qwen3.8 27B

  The strongest all-rounder for a 24GB card, and coding agents. Apache 2.0.

  ollama run qwen3.8:27b

• Fits 64–128GB

  gpt-oss-120b

  Close to a hosted experience on a 128GB machine. Apache 2.0.

  ollama run gpt-oss:120b

27 models, with licenses, sizes and sources, in the model finder. The big families are graded on the open-models shelf.

The apps that run them

• Easiest first installLM StudioThe easiest start: a polished app that finds, downloads and runs models, with a local server when you want one. The app itself is not open source.
• The standard base to build onOllamaThe default engine most other local apps plug into: one command to run a model, and an OpenAI-compatible API on your machine.
• Open-source desktop chatJanThe open-source answer to LM Studio: a clean chat app that runs models offline, with a local API and MCP tools.
• Your own ChatGPT for the houseOpen WebUIThe most complete front end: accounts, documents, memory and web search on top of Ollama or any OpenAI-style API.

Nine apps compared, with difficulty and license, in local AI apps. Not sure your computer is up to it? Read the hardware guide. Straight answers to the common questions, like Ollama vs LM Studio or the best local coding model, are in comparisons.

Start tonight

1. Find your computer's memory, then check what fits.
2. Install LM Studio or Ollama and run your first model: fifteen minutes, step by step.
3. Put your instructions in one file so every model you try already knows you.

Then keep going: the learning path takes you to a setup the whole house can use.

Path A: on your own computer

• Good for: privacy, working offline, and people who already own a 16GB+ Mac or a gaming PC.
• Cost: $0 on hardware you already own. Buying new hardware just to replace a $20 subscription takes about five years to pay back, and 2026 hardware prices are inflated.
• Quality: what fits on a normal computer is a clear step down from ChatGPT or Claude. Fine for drafts, summaries and chat.
• Who sees your data: no one. Nothing leaves your machine. Watch for “cloud” models inside local apps: those run on someone else's servers.
• Start with: LM Studio or Ollama. Then check Will it run on my machine?
• Buying hardware? Compare the best hardware to run an LLM locally.

Path B: a cheap monthly plan

• Good for: people leaving a pricey ChatGPT or Claude plan who want the best open models without new hardware.
• Quality: the best open models, about four months behind the best paid models.
• Who sees your data: the plan's company. Check below before you sign up.
• Rather rent a GPU by the hour? Compare the best GPU cloud for LLM hosting.

Path C: a router (one key, any model)

• Good for: trying many models and paying only for what you use.
• Cost: OpenRouter is free for 50 requests a day on its free models. A one-time $10 top-up raises that to 1,000 a day and unlocks every paid model. Its fee is 5.5%.
• Who sees your data: OpenRouter doesn't log prompts by default, and you can exclude providers that do. If a provider fails, it can switch to another one; pin a provider if you need to know exactly who runs your request.
• Other routers: compare OpenRouter alternatives.

Same model, different landlord

Where you run DeepSeek, GLM or Kimi decides where your words go.

Keeping your data out of China costs almost nothing: the same models on US hosts are priced about the same as the labs' own.

The model can change. Your continuity should stay.

Run your own AI for long and you will switch models every few months. What should not reset is everything the AI knows about you and your work. That is what this site is about: the continuity layer, the part that stays when the model, the app or the company changes. Start with one file any model can read, and when you outgrow it, build your own continuity layer.