Open vs
closed.
There are two fundamentally different ways an AI model can exist in the world: locked inside a company's servers and rented by the token, or published for anyone on earth to download and run. That single distinction shapes cost, privacy, control, and who holds the power in AI.
This is a different question from frontier vs local. That one is about where a model runs. This one is about who controls the weights — whether the actual model file is a secret or a public download. The two often line up, but they're not the same axis, and understanding both is what lets you make smart AI decisions instead of just picking a brand.
Renting a model you can't see inside
A closed-source model is the kind most people use: GPT, Claude, Gemini. The company that built it keeps the weights private and lets you use the model only through their API or app. They control everything — the price, who's allowed access, when it changes, and when it's retired.
You get the frontier of capability with zero infrastructure to run. What you give up is control and transparency: you can't download it, can't run it offline, can't fully inspect how it behaves, and your data leaves your building on every request. You're a tenant, not an owner.
A model you can download and keep
An open-source (more precisely, open-weight) model has its trained weights published. Llama, Qwen, DeepSeek, Mistral, Kimi — you can download the actual model file, run it on your own hardware, inspect it, and fine-tune it for your own use. No one can revoke your copy or change the price on you.
Open weights means you get the finished model to run. Fully open goes further — the company also publishes the training data, code, and techniques (the whole recipe). Most "open" models today are open-weight; a smaller number are fully open. Both let you run the model yourself.
Why give away something this valuable?
It seems backwards — why spend a fortune building a world-class model and then hand it out for free? Because the model isn't always where the company plans to make its money. Open-sourcing a layer you want to commoditize lets you compete harder in the layers where you profit. A few reasons a company does it:
- Become the standard. If everyone builds on your model, you shape the ecosystem and the tech beneath it.
- Pressure competitors. Free, capable models squeeze the margins of companies selling closed ones — sometimes called a "scorched-earth" move.
- Sell the shovels. Whoever makes chips, cloud, or tooling wins when more people run more models, no matter whose model it is.
It's the same playbook as Linux (runs much of the internet), Android (most of the world's phones), Chromium (the engine behind most browsers), and React (how much of the web's front-end is built). Give away the commodity layer; win everywhere around it.
Open doesn't always mean cheaper
Downloading the weights is free. Running a frontier-scale model is not: you need serious hardware and the electricity to feed it, plus the work of keeping it running. That's why most open models are also offered through paid APIs — you rent them by the token just like a closed model, often at a lower price. "Open" buys you freedom and control, not automatically a lower bill. Whether it's actually cheaper comes down to the real cost per task, which is its own topic.
Optionality is the real prize
For a business, the open ecosystem is a gift even if you never download a model yourself. It means you're not locked to one vendor's price and roadmap; there's competitive pressure driving costs down; and when privacy or control matters, you have the option to run a capable model entirely in-house. The strongest setups stay flexible — a closed frontier model for the hardest work, open models for everything private, high-volume, or cost-sensitive.
That flexibility is exactly what we build for clients: private AI systems on hardware you own, so you get the control of open without having to become an infrastructure company. It's the same reason our own stack runs open models on a rack in the office.
Common questions
What is the difference between open-source and closed-source AI?
A closed-source model (GPT, Claude, Gemini) is controlled entirely by the lab that built it — you reach it only through their API, and they set price, access, and rules. An open-source / open-weight model (Llama, Qwen, DeepSeek) has its weights published, so anyone can download it, run it on their own hardware, inspect it, and fine-tune it.
Why do companies give away AI models for free?
A free, strong model drives adoption and becomes a standard, giving the releasing company influence and pulling demand toward the parts of the stack where they do make money (chips, cloud, tooling) — while squeezing competitors' margins. It's the same commoditize-your-complement strategy behind Linux, Android, Chromium, and React.
Is open-source AI actually free?
The weights are free to download, but running a frontier-scale model yourself needs expensive hardware and electricity, plus the work of operating it — and most open models are also sold through paid APIs. Open means freedom and control, not necessarily zero cost.
What are examples of open and closed AI models?
Closed/API-only: OpenAI's GPT, Anthropic's Claude, Google's Gemini. Open-weight: Meta's Llama, Alibaba's Qwen, DeepSeek, Mistral, and Moonshot's Kimi. Open models have closed much of the capability gap with the leading closed ones.