Open or closed AI models: what each one means and which suits you

When you use ChatGPT, Claude or Gemini on their websites, you are using a model that lives on the servers of the company that built it. You cannot download it or see how it works inside: you just send text and get answers. But there is another category of models that you can download, run on your own device and even adapt. Understanding the difference helps you choose better, especially if you work with sensitive data or large volumes.

What a closed model is

A closed (or proprietary) model is only available as a service. You access it through an app or an API, you pay by subscription or by usage, and the company controls everything else: which servers it runs on, which version you use, when it changes and what it does with your data under its terms.

The advantages are clear: they are usually the most capable models available, you need no special hardware and you can start using them in a minute. So are the drawbacks: you depend on a provider, your text leaves your device and the cost grows with use.

What an open-weights model is

A language model is, in essence, a huge file with billions of numbers (the "weights") that were adjusted during training. When a company publishes that file for anyone to download, it is called an open-weights model. With it you can run the model on your computer or server, without going through the company that trained it.

Well-known open-weights families include Meta's Llama, Mistral, Google's Gemma, Alibaba's Qwen and DeepSeek. They are usually downloaded from Hugging Face, the most widely used model repository.

Open weights is not the same as open source

This is the most common confusion. Being able to download the weights does not mean the model is open source in the strict sense.

• The licence may attach conditions. Llama's community licence, for example, allows commercial use, but requires companies with more than 700 million monthly active users to request a separate licence from Meta, and includes an acceptable use policy. Other models are released under permissive licences such as Apache 2.0, which impose very few restrictions.
• What is behind it is missing. The full training data and all the training code are almost never published. You can use the model, but you cannot reproduce it from scratch.

That is why the Open Source Initiative, the organisation that maintains the definition of open source in software, published its Open Source AI Definition 1.0 in October 2024. Under it, an open-source AI system must allow you to use, study, modify and share it, which requires access to the weights, the code and sufficient information about the training data. Many models marketed as "open" do not meet that definition.

The practical rule: before using an open-weights model in a product or a business, read its licence. It is a short document and it says what you can and cannot do.

What you need to run a model locally

Running a model on your own device no longer requires systems expertise. Tools such as Ollama or LM Studio let you download a model and chat with it in a few minutes, without coding. What does matter is the hardware:

• Memory. The model has to fit in the graphics card's memory or in the device's unified memory. As a rough guide, a 7 or 8 billion parameter model quantised to 4 bits takes about 4 or 5 GB, and runs on many recent laptops with 16 GB. Models with 70 billion parameters need far more and are out of reach for a typical laptop.
• Quantisation. This technique shrinks the model by storing each weight with less precision (for example 4 bits instead of 16). The model takes up less space and runs faster, at the cost of some quality. GGUF files, the format used by llama.cpp and leveraged by Ollama under the hood, are the most common way to distribute quantised models.
• Small models. Small language models (SLMs) are designed precisely for modest hardware. They do not compete with large models on complex reasoning, but they handle bounded tasks such as classifying, summarising or extracting data well.

Comparison on what matters

• Privacy. A local model sends your text nowhere. With a closed service, your data is processed on someone else's servers and you depend on their terms (we explain this in What never to paste into an AI chatbot).
• Quality. The most advanced closed models are usually ahead on hard tasks. The gap with the best open models has narrowed, but models that fit on a laptop are still behind.
• Cost. In the cloud you pay per token or by subscription. Locally you pay for the hardware and the electricity, but each additional query costs nothing. For high volumes and repetitive tasks, local can be cheaper; for occasional use, it almost never is.
• Control. A downloaded model does not change unless you decide it should. A closed service can update or retire a model and alter results you took for granted.
• Effort. A closed service works with no maintenance. Your own setup requires installing, updating and monitoring.

How to decide

Ask yourself three questions, in this order:

1. Can the data leave your device? If you work with information that must not leave (health data, customer data, confidential documents), a local model, or an enterprise service with contractual guarantees, is the sensible option.
2. How much will you use it? For occasional use, a closed service is more convenient and cheaper. If you will process thousands of documents a month with a repetitive task, it is worth calculating the cost of doing it locally.
3. What quality do you need? For complex reasoning, advanced programming or sensitive writing, large models still make the difference. For classifying or summarising, a well-chosen small model is usually enough.

It is not an either-or decision: many people use a closed model for demanding work and a local one for sensitive or repetitive tasks. To compare specific candidates without being swayed by rankings, follow the method in How to choose an AI model without looking at the leaderboard, and if you want to set out the decision in writing, the Decide with criteria template helps.

Keep reading

• How to choose an AI model without looking at the leaderboard: your own tests with your real tasks.
• What is a large language model (LLM): what parameters and weights are.
• What never to paste into an AI chatbot: which data to protect.
• Tokens and the context window: how usage and cost are measured.
• In the glossary: open weights, quantisation, GGUF and SLM.

Sources

• Open Source Initiative (2024). The Open Source AI Definition 1.0.
• Meta. Llama 3.2 Community License Agreement. Hugging Face.
• llama.cpp. Repository and GGUF format documentation.
• Ollama. Ollama documentation.