You’re writing a novel with an AI assistant, and a thought nags at you: where is my manuscript going? When you type a scene into ChatGPT or Claude, your words travel to company servers, get processed, and come back as a response. You’re trusting someone else with your unpublished work. For some authors, that’s fine. For others, it’s like handing your diary to a stranger and hoping they don’t peek.
Open-source models offer a different arrangement. Instead of accessing AI through a company’s locked door, you download the model itself and run it on your own computer. Your writing never leaves your machine. Nobody sees it, nobody stores it, nobody trains on it. The AI lives on your desk, not in someone else’s cloud.
But “open source” in AI doesn’t always mean what you’d expect, and the fight over this term reveals a lot about where the technology is heading.
A Term with Decades of Baggage
“Open source” meant something specific long before AI models existed. The concept traces back to 1983, when Richard Stallman, a programmer at MIT, launched the GNU Project after he couldn’t modify the software on a printer in his lab. His philosophy was radical: software should be free to use, study, modify, and share. He founded the Free Software Foundation and created the legal framework that would make this vision possible.
In 1991, a Finnish student named Linus Torvalds used Stallman’s tools to build the Linux kernel, which now runs most of the internet’s servers. The label “open source” was coined in 1998 by Christine Peterson as a more business-friendly rebrand of Stallman’s “free software.” (Stallman still rejects the new name. He’s nothing if not consistent.)
For decades, the principle was straightforward: you release the source code under a license that lets anyone use and modify it. Then AI models arrived, and the definition started to stretch.
What Gets Released (and What Doesn’t)
A large language model isn’t just code. It has distinct components: the architecture (the neural network blueprint), the weights (billions of learned numerical values where the “intelligence” actually lives), the training code (the scripts used to build the model), and the training data (the text the model learned from).
A truly open-source model would release all four under a license that lets anyone use, modify, and share them freely. In practice, almost nobody does this. What most companies actually release is the weights and architecture, letting you download and run the model, while keeping the training data and training code private. This is more accurately called “open weight.”
The distinction matters. Meta calls its Llama models “open source.” The Open Source Initiative, the organization that literally defines the term, says they’re not. Meta’s license restricts who can use the models (companies with over 700 million users need special permission) and how (you can’t use Llama’s output to train competing models). The Free Software Foundation classified Llama’s license as “nonfree” in early 2025.
This isn’t pedantry. When an author hears “open source,” they might reasonably assume no restrictions apply. “Open weight” is the more honest label for most models people casually call open source. That said, genuinely open-source AI models do exist. EleutherAI’s Pythia, the Allen Institute’s OLMo, and the BigScience project’s BLOOM all released everything under permissive licenses. They get less attention than Meta’s flashier launches, but they’re the ones actually living up to the name.
A Short, Dramatic History
The timeline of open AI models is compressed and, at one pivotal moment, involves 4chan.
In August 2022, Stability AI released Stable Diffusion, an image generation model anyone could download and run on their own hardware. It was a proof of concept: powerful generative AI didn’t have to live behind an API.
Then came February 2023. Meta released LLaMA for academic research, with access limited to approved applicants. That lasted about a week. Someone uploaded the model weights to 4chan, and within days they’d spread across the internet. The leak was unauthorized, but spectacularly consequential. It sparked an explosion of community projects (Stanford’s Alpaca, Vicuna, dozens of others) that might never have existed otherwise.
Meta read the room. By July 2023, the company officially released LLaMA 2 with a broadly permissive license, followed by LLaMA 3 in 2024 and LLaMA 4 in 2025. Mistral AI, founded in Paris by former Meta and Google researchers, built its identity around competitive open-weight models. Google released Gemma. China’s DeepSeek lab made waves with models rivaling the best closed offerings. In three years, the gap between what you can access freely and what sits behind a paywall shrank dramatically.
Why This Matters for Your Writing Life
Open-source models aren’t just a tech industry debate. They show up in the writing tools you use every day.
Privacy you can actually verify. When you run a model on your own computer, your writing stays private in the most literal sense. Not “we promise we won’t look” private. Physically, mathematically, your-words-never-leave-your-machine private. For authors working on unpublished manuscripts or sensitive material, this is a meaningful difference.
Tools that connect to open models. NovelAI’s text model, Erato, is built on Meta’s Llama 3, fine-tuned on storytelling data. NovelCrafter lets you connect directly to models running on your own machine through tools like Ollama or LM Studio. KoboldCpp is an open-source application built specifically for creative writing with local models. These are real writing tools that thousands of authors use daily.
Fine-tuning becomes possible. With a closed model like GPT-4, you can’t train the AI on your previous novels to learn your voice. With an open model, you can. Feed it your backlist and it generates prose that sounds like you. Train it on your worldbuilding documents and it maintains consistency across your series. This is exactly what NovelAI’s AI Modules and NovelCrafter’s Fine-Tune Dataset Editor are built for, and it only works because the underlying models are open.
Cost drops significantly. Closed models charge per use, typically by the token. Open models, once downloaded, cost nothing to run beyond electricity. Running a smaller model on a modern laptop with 16 GB of RAM is entirely feasible. No monthly fees, no usage caps.
The tradeoff is honest. The most capable models (GPT-4, Claude, Gemini) are still closed. Open models have gotten remarkably good, but the absolute cutting edge remains behind the paywall for now. Running models locally also requires some technical comfort, though tools like LM Studio have made it as simple as downloading an app and clicking a button. For many writing tasks, an open model on your own hardware is already more than good enough.
Understanding this landscape helps you make informed choices about which AI tools deserve your trust, your money, and your manuscripts. Do you want to rent access to someone else’s AI, or do you want to own the tool outright? Both are valid answers. The important thing is knowing you have the choice.