Open-weight model explained simply: how open AI models differ from open source and ChatGPT

What “weights” are and what is open about them

A language model learns during training by adjusting billions of numerical values over and over again. These numbers are called weights or parameters. The Open Source Initiative (OSI) describes them as the final values of a trained neural network, which determine how the model processes input and produces output. You can think of them as the model's stored know-how.

With an open-weight model, the maker publishes these weights as a file. Anyone with enough computing power can download them, run the model on their own hardware and adapt it further.

The opposite is a closed model. Its weights stay with the provider, and you only use it through the provider's app or programming interface. That is the case for the models behind ChatGPT, Gemini and Claude, and since April 2026 also for Meta's top model Muse Spark, which was released without downloadable weights.

Open weight is not the same as open source

Many makers and media outlets say “open source” when they actually mean “open weight”. In 2024, the OSI published its own definition of open-source AI (version 1.0). Under it, genuine open-source AI includes not only the weights but also the complete code used for training and information on the training data detailed enough for experts to build a substantially equivalent system.

That is exactly what most open-weight models lack. The OSI criticises that their development and possible biases then cannot be fully traced. DeepSeek, for example, only says on its model page that V4 was trained on more than 32 trillion text units (tokens), not which ones.

Then there is the licence, which sets out what you may do with the weights. Apache 2.0 and MIT are very permissive. Other licences impose limits: according to the Llama 4 acceptable use policy, individuals domiciled in the EU and companies with their principal place of business in the EU are not granted usage rights for the multimodal Llama 4 models. According to Meta, people who only use an app that contains such a model are not affected.

Well-known open models at a glance

OpenAI: in August 2025, OpenAI released gpt-oss-120b and gpt-oss-20b under the Apache 2.0 licence. According to OpenAI, the smaller model needs only 16 GB of memory, while the larger one runs on a single 80 GB graphics card.

Meta: with Llama, Meta had one of the most downloaded model families with open weights. After the closed Muse Spark, Meta released an open model again in August 2026 with Muse Glimmer: 30 billion parameters, Apache 2.0 licence and, according to Meta, able to run on a Mac or PC with one powerful graphics card.

Mistral: the French company's model overview lists, among others, Mistral Small 4 and Mistral Large 3 under Apache 2.0 and Mistral Medium 3.5 under a modified MIT licence.

DeepSeek: according to the model page, DeepSeek released the V4 models as a preview version, under the MIT licence. V4-Pro has 1.6 trillion parameters, V4-Flash 284 billion.

Qwen: Alibaba releases many Qwen models with open weights. In August 2026, the team announced that it would for the first time also release the weights of a model in its largest Max class.

What lies behind DeepSeek and Qwen, and how these models fare on privacy and censorship, is covered in detail in our article on Chinese AI models.

What it means for privacy and your own computer

Open weights are the precondition for running a model locally, meaning your input never leaves the device. How that works in practice and which computer is enough is explained in our article on local AI models.

Our assessment: open weight does not automatically mean private. If you use the same model through the maker's app or website, your input still ends up on its servers. It only becomes private when the model really runs on your machine or with a service provider you trust.

Also, not every open model fits on a home computer. Models with hundreds of billions or more than a trillion parameters, such as DeepSeek V4, need server hardware. For home use, small models such as gpt-oss-20b or Muse Glimmer are more realistic.

For businesses, the licence is decisive: Apache 2.0 and MIT generally allow commercial use as well, while for other licences, such as Llama 4's, you should read the terms carefully first.