Open-Source vs. Proprietary AI Models

"Open-source AI" gets thrown around loosely enough that it's worth being precise about what it actually means, especially since it's a completely different axis from the local-versus-cloud distinction covered elsewhere in this help center. Local vs. cloud is about where a model runs. Open-source vs. proprietary is about whether you have access to the model itself, and under what terms. The two questions happen to be related, since you can only self-host a model whose weights someone has actually published, but they're not the same question.

Three Categories, Not Two

Most discussion collapses this into a binary, but there are really three distinct tiers:

Closed, API-only models. Claude, GPT, and Gemini fall here. The trained weights are never published. You send a request to the provider's servers and get a response back; there is no version of "download the model" available to you at any price.
Open-weight models. The trained parameters are published and downloadable, so you can run the model yourself, but the training data, training code, and full development pipeline usually aren't disclosed. This is where the overwhelming majority of models marketed as "open source" actually sit, including Llama, DeepSeek, and Qwen.
Fully open-source models. Weights, training code, and training data are all public, meeting the Open Source Initiative's Open Source AI Definition. At frontier scale, this is genuinely rare; models like OLMo and Pythia qualify, but none of the widely used consumer-facing models do.

The distinction between the second and third tiers isn't pedantic. "Open-weight" gives you the finished product; "open-source," in the strict sense, gives you the recipe as well. Almost everything people casually call open-source AI is open-weight.

Why This Determines What You Can Self-Host

This is the direct link back to local vs. cloud AI: local hosting is only possible for models whose weights have been released. That's exactly why DeepSeek can run on your own server through Ollama, and why Claude Code defaults to calling Anthropic's cloud API even though the Claude Code application itself is installed on your VPS or VDS. Anthropic has never published Claude's weights, so there's no local version to run in the first place, regardless of how the surrounding app is deployed.

Licensing in Practice

Model Family	Category	Practical Terms
DeepSeek (R1, V3)	Open-weight	Released under the MIT License, one of the most permissive licenses that exists. Commercial use, modification, fine-tuning, and redistribution are all allowed with essentially no restriction beyond preserving the copyright notice. Some older DeepSeek artifacts sit under a separate, more restrictive custom license, so it's worth checking the specific release rather than assuming MIT applies universally.
Llama	Open-weight	Governed by Meta's Community License, which is not OSI-approved. It restricts use above a monthly active user threshold, requires derivative models to include "Llama" in the name, and explicitly prohibits using Llama's output to train competing models. Meta calls this open source in its own marketing; the Open Source Initiative and Free Software Foundation both dispute that characterization.
Claude, GPT, Gemini	Closed	No published weights under any license. Access exists only through the provider's API or products built on top of it. There is no self-hosting path for these models, full stop.

What This Actually Changes for You

Inspectability. With an open-weight model, you can examine the architecture, run it in an isolated environment, and verify what it's doing with your data, since nothing leaves your server. With a closed model, you're trusting the provider's stated policies, since the internals aren't available to check.
Fine-tuning and customization. Open-weight models can be fine-tuned on your own data if you have the hardware for it. Closed models generally only offer this through the provider's own managed fine-tuning service, if they offer it at all.
License compliance is a real, ongoing obligation. Deploying Llama commercially at scale isn't the same legal situation as deploying DeepSeek commercially at scale. If your use case is more than personal experimentation, the specific license terms are worth reading rather than assuming "open" means "unrestricted."
Reliability and support trade off against control. A closed model's behavior, uptime, and safety tuning are the provider's responsibility. An open-weight model you're self-hosting puts all of that on you, in exchange for full control over the deployment.

Summary

Open-source, open-weight, and closed are three distinct categories, not two, and almost every model marketed as open source, including DeepSeek and Llama, is technically open-weight: the trained parameters are public, but the training data and code generally aren't. This distinction is what makes self-hosting possible at all, since only models with published weights can be run locally in the first place, and it's why an app like Claude Code still depends on a cloud API regardless of where the app itself is installed. The license attached to an open-weight model, whether that's DeepSeek's permissive MIT terms or Llama's more restrictive Community License, determines what you're actually allowed to do with it, and that's worth checking directly rather than assuming from the "open" label alone.