Open vs Closed AI Models: Open Source LLMs Explained

A closed AI model is one you can only use through the company that makes it, usually via an app or an API. An open model is one whose weights, the trained numbers that make up the model, are published so you can download and run it yourself. Many models called “open source LLMs” are more precisely “open-weight,” and their licenses vary a lot.

That nuance matters when you choose a model. Here is what the labels mean and how the trade-offs play out.

What “open” and “closed” actually refer to

A language model is, at its core, a very large set of learned parameters (the weights) plus the code that runs them. Whether a model is open or closed mostly comes down to who can access those weights.

Closed models

With a closed (or proprietary) model, the weights stay on the provider’s servers. You send your prompt over the internet, the provider runs the model, and you get the answer back. Claude, ChatGPT, and Gemini are well-known examples of this approach.

You never see the model itself. You use it as a service, under the provider’s terms.

Open-weight models

With an open-weight model, the provider publishes the weights for download. You can run the model on your own computer or server, inspect how it responds, and often adapt it with further training. Meta’s Llama family and many of Mistral’s models are commonly cited examples, and there are many others from companies and research groups.

Tools like Ollama make running these models on a laptop fairly approachable.

Open source vs open-weight: the nuance

In traditional software, “open source” means you get the source code under a license that lets you use, study, change, and share it, generally without restricting who can use it or for what.

AI models complicate that definition. To fully reproduce a model, you would need more than the weights:

• Weights: the trained model you can run.
• Training code: how the model was trained.
• Training data: what it learned from, or at least a detailed description of it.
• License: the legal terms for using all of the above.

Many popular “open” models publish weights and some code, but not the full training data. That is why many people prefer the term open-weight for them. Models that release weights, code, and data details under a permissive license are closer to open source in the classic sense.

Licenses vary more than you might expect

Two models that both call themselves open can come with very different terms. Some common patterns:

• Permissive licenses that allow broad commercial use with few conditions.
• Custom model licenses that allow most uses but add restrictions, such as acceptable-use rules or conditions for very large companies.
• Research or non-commercial licenses that let you experiment but not build a paid product.

Always read the license of the specific model before using it in a product. “Open” in a headline does not tell you what you are allowed to do.

The trade-offs

Neither approach wins everywhere. It depends on what you value.

	Closed models	Open-weight models
Access	App or API from the provider	Download and run yourself, or use a host
Capability	Often the strongest general models	Wide range; strong options exist, especially for focused tasks
Privacy	Data goes to the provider, under their policies	Can stay entirely on your own hardware
Cost model	Pay per use or subscription	Your hardware and time, or a hosting bill
Setup	Minimal	More effort, from easy to substantial
Control	Provider decides updates and changes	You choose the exact version and keep it
Customization	Limited to what the provider offers	Can be fine-tuned and modified, license permitting

Why people choose closed models

• Top capability with no setup. For hard reasoning, long and complex coding tasks, or broad general knowledge, the strongest available models tend to be closed.
• Managed infrastructure. No GPUs to buy, no servers to maintain.
• Built-in extras such as safety systems, tool support, and large context windows.

Why people choose open-weight models

• Privacy and data control. Running a model locally means your prompts and files never leave your machine.
• Predictable behavior. The model won’t change underneath you unless you change it.
• Offline use and freedom from rate limits.
• Customization. You can fine-tune for a niche task; our explainer on RAG vs fine-tuning covers when that makes sense.
• Cost at scale for some workloads, once you own suitable hardware.

The hardware reality

Open-weight models come in many sizes. Small ones run on an ordinary laptop; the largest need serious GPUs. On typical consumer hardware, smaller local models are noticeably slower and less capable than the top closed models, especially for complex, multi-step work. Our comparison of local vs cloud AI models goes deeper.

Which should you use?

A practical rule of thumb:

1. Start with a closed model if you want the best results quickly and your data can be shared with a provider under their terms.
2. Choose an open-weight model if your data must stay on your hardware, you need offline use, or you want full control over the exact model.
3. Mix them when you can. Use a strong closed model for hard tasks and a local model for simple, private, or high-volume ones.

Mixing is increasingly common in agent setups. In BotQuarters, for example, one team can include teammates running on cloud services like Claude Code or Codex alongside teammates running local models through Ollama.

Common misconceptions

• “Open means free.” The weights may be free to download, but running a model costs hardware, electricity, or hosting.
• “Open means anything goes.” The license sets the rules, and some are restrictive.
• “Closed means my data trains the model.” Policies differ by provider and plan. Read the terms for the service you use.
• “Open models are always worse.” The strongest models are often closed, but open-weight models can be excellent for focused tasks, and the right small model can beat a general one on a narrow job.

FAQ

What is the difference between open source and open-weight AI?

Open-weight means the trained model weights are published so you can run them. Open source, in the stricter sense, would also include the training code, information about the training data, and a license that allows broad use and modification. Many popular “open source LLMs” are really open-weight.

Are open source LLMs free for commercial use?

Some are, some are not. Licenses range from permissive to restricted or non-commercial, and some add conditions for large companies or specific uses. Check the license of the exact model before building a product on it.

Are closed AI models better than open models?

The most capable general-purpose models have often been closed, especially for complex reasoning and coding. Open-weight models can still be the better choice when privacy, offline use, cost control, or customization matter more than peak capability.

Can I run an open source LLM on my laptop?

Yes, smaller open-weight models run on many modern laptops using tools such as Ollama. Expect them to be slower and less capable than large cloud models, and choose a model size that fits your computer’s memory.