Is an open-weight model the same as an open source model?

No. Open weights means the trained weights can be downloaded. The Open Source AI Definition 1.0 also asks for sufficiently detailed information about the training data and the complete source code used to train and run the system, all under OSI-approved terms. A model with downloadable weights but no such data information and code is open-weight.

Can I run an LLM on my own computer?

If the weights are downloadable, yes, provided the machine has enough memory to load them. Quantized versions need less memory, and tools such as llama.cpp run them on a wide range of hardware. The memory a specific variant needs is listed on its model card.

Does downloading a model keep my data private?

When the model runs on your own machine, your prompts are processed there and are not sent to a vendor. If you instead use a hosted service that runs an open-weight model for you, your prompts go to that service, under its terms.

Is Llama or ChatGPT better?

They are different kinds of thing, so this page does not rank them. Llama is a family of models whose weights can be downloaded under the Llama license. ChatGPT is a chat product from OpenAI that runs on OpenAI's servers. The question to ask is which kind of access fits what you need.

Can I use an open-weight model for a business?

It depends on the license. The Apache License 2.0 grants a no-charge, royalty-free copyright license. The Llama 3.1 Community License Agreement adds conditions, such as a license request above 700 million monthly active users and naming and attribution rules. Read the license of the exact model you use.

LLM · MODEL TYPES

Open-Weight vs Closed LLMs: What "Open Source" Really Means

An open-weight LLM is one whose trained weights you can download and run yourself. A closed LLM stays on the vendor's servers and you reach it through their app or API. "Open source" is a stricter label than either, and most models people call open source do not meet it.

One-pager titled "Open weights" with the subtitle "You can download the numbers, not the recipe": a left panel "You download" holds three chips, Weights (highlighted in amber, "the learned numbers"), Configuration and Tokenizer; a right dashed panel "Stays with the maker" holds two dashed chips, Training data and Training code, with the amber line "Data + code = the recipe"; a REMEMBER band reads "Numbers, not the recipe" under a Weights chip, "Open source needs all three" under Weights, Data info and Code chips, and "The license sets the rules" under a License chip. — One-pager: what an open-weight download contains, and what stays with the maker.

What is actually in a model download

A trained LLM is a set of files. The largest one holds the weights: the learned numbers that make up the neural network. Next to it sit a small configuration file, which describes the shape of the network, and a tokenizer, the lookup table that turns text into token IDs.

The Mistral-7B-v0.3 model card on Hugging Face is a plain example. It lists three files to download: consolidated.safetensors, params.json and tokenizer.model.v3. Safetensors is a format for storing tensors, the arrays of numbers a network is made of, safely and quickly.

Those files are enough to run the model. They are not enough to recreate it. The data it learned from and the code that did the training are separate things, and a model card can leave them out. As of 2026-10-11, the Mistral-7B-v0.3 card says nothing about its training data.

Open weights, open source and closed are three different things

Open weights means the weight files can be downloaded. Closed means the model runs only on the vendor's servers and you reach it through their app or API. The weights are never handed over.

Open source has a stricter, published definition. The Open Source Initiative's Open Source AI Definition 1.0 asks for the freedom to use, study, modify and share a system. It says the preferred form for making modifications has three parts: sufficiently detailed information about the training data, the complete source code used to train and run the system, and the model parameters. Each is to be made available under OSI-approved terms.

The OSI's FAQ adds that training data which cannot legally be shared must still be described in detail. So a model that ships only its weights is open-weight, not open source. This page uses "open-weight" for those, even where headlines say "open source LLM".

The license sets what you may do with the files

Downloadable does not mean unrestricted. The terms sit in a license, and they differ from model to model. Two published texts, read as of 2026-10-11, show the contrast.

The Mistral-7B-v0.3 card states "apache-2.0". The Apache License 2.0 grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license". It requires that recipients get a copy of the license, and it does not grant permission to use the licensor's trade names or trademarks.

The Llama 3.1 Community License Agreement (release date July 23, 2024) is different in kind. If your products had more than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which may grant it at its sole discretion. If you distribute the materials or a product containing them, you must display "Built with Llama". If you use the materials or their outputs to create, train, fine-tune or improve an AI model that you distribute, its name must begin with "Llama". Use must also follow an Acceptable Use Policy.

Read the license for the exact model and version you download. A model card's license line is the starting point, not the whole answer.

What the difference changes in practice

With a closed model, your prompt travels to the vendor's servers, the model runs there, and only the reply comes back. What happens to the data is set by the vendor's terms. As of 2026-10-11, OpenAI's API documentation says data sent to its API is not used to train its models unless you opt in, and that abuse-monitoring logs are retained for up to 30 days. Other vendors have their own terms.

With an open-weight model, you can load the weights into software on a machine you control, so the prompt is processed there. That machine can also be a rented one, in which case the data goes to whoever runs it. The practical limit is memory, because the weights have to be loaded to run.

Quantization stores the weights at lower precision to cut that memory, at some cost in accuracy. Projects such as llama.cpp are built to run quantized models on a wide range of hardware. How much memory a given variant needs is stated on that variant's own model card.

Holding the weights also means you can keep training them. That is fine-tuning, and a LoRA is a lighter way to do it, subject to the license.

What open weights still do not give you

The training data stays with the maker unless the maker publishes it. The training code may or may not be published. Without both, you cannot reproduce the training run that produced the weights.

A download is a finished snapshot. You can change its behavior by training it further, but you cannot ask for the original training to be done differently.

None of this says one kind of model is better than the other. It describes what you are given: files you hold, or an interface someone else operates.

FAQ

Is an open-weight model the same as an open source model?: No. Open weights means the trained weights can be downloaded. The Open Source AI Definition 1.0 also asks for sufficiently detailed information about the training data and the complete source code used to train and run the system, all under OSI-approved terms. A model with downloadable weights but no such data information and code is open-weight.
Can I run an LLM on my own computer?: If the weights are downloadable, yes, provided the machine has enough memory to load them. Quantized versions need less memory, and tools such as llama.cpp run them on a wide range of hardware. The memory a specific variant needs is listed on its model card.
Does downloading a model keep my data private?: When the model runs on your own machine, your prompts are processed there and are not sent to a vendor. If you instead use a hosted service that runs an open-weight model for you, your prompts go to that service, under its terms.
Is Llama or ChatGPT better?: They are different kinds of thing, so this page does not rank them. Llama is a family of models whose weights can be downloaded under the Llama license. ChatGPT is a chat product from OpenAI that runs on OpenAI's servers. The question to ask is which kind of access fits what you need.
Can I use an open-weight model for a business?: It depends on the license. The Apache License 2.0 grants a no-charge, royalty-free copyright license. The Llama 3.1 Community License Agreement adds conditions, such as a license request above 700 million monthly active users and naming and attribution rules. Read the license of the exact model you use.

Sources

Last updated 2026-10-11