open-weight model

basicwidespread

An AI model whose trained parameters (weights) are published by its creator, allowing anyone to download, run on their own machine, and further fine-tune it according to the license terms.

In plain terms

An AI model's 'knowledge' is stored in vast numbers of so-called weights. Large models can contain billions of these. In an open-weight model, the developer releases this set of numbers as a downloadable file. From then on, you do not need to send your questions to their server: the model can run on your machine or your company's server.

It is like when the pastry chef gives you the finished cake instead of the recipe. You can take it home and re-decorate it, but you still do not know exactly what it is made of or how it was prepared, and you cannot bake it again. You receive the weights, but usually not the training data or the entire training process.

In contrast, closed models are accessible only through the manufacturer's service: the model stays with them, and you can only ask questions. Open-weight models are advantageous if it is important that your data does not leave your premises, that you do not depend on a single provider, or that you customize the model for your own tasks.

How it works

The developer company trains the model, which demands huge computing capacity and vast data.
It uploads the finished weights as a file to a public model repository and attaches a license that sets what you may use them for.
You download the file and run it on your own machine or server using an executable. Smaller models can run on a good laptop, while the largest require a server equipped with multiple powerful graphics cards.
If you wish, you can further train it (fine-tune) with your own data, compress it, or embed it in your own application, such as behind an agent.
From now on, you are largely responsible for execution, security, and updates. The model creator also bears certain obligations, for example under the EU AI Act, but you decide what to use the model for, and you are responsible for that.

An everyday example

A law firm does not want client contracts to go to an external cloud provider. It downloads an open-weight model, runs it on its own server, and uses it to summarize documents, so the text never leaves the office.

Who is affected

developerYou can freely download, fine-tune, and integrate it into your own product without paying a provider for every request. → Code

entrepreneur, executiveYou can process sensitive business data in-house and will not be at the mercy of a single vendor's prices or decisions.

lawyer, citizenLicensing rules and the EU AI Act determine what the model can be used for and who is responsible. → Law

parent, student, teacherFor free, even without internet, you can experiment with an AI model on your own machine and learn its operation up close. → Education & Family

A common misconception

MythAn open-weight model is the same as open-source and can be freely used for anything.

In realityPublishing weights does not mean that the training data and training code are also public. Licenses also differ: some allow almost everything, while others restrict, for example, business applications with large user numbers or certain use cases. Always read the license before building on it.

Limits and risks

The weights do not reveal what data the model was trained on, making it difficult to check its biases and copyright risks.
Running large models requires expensive hardware and expertise, meaning free download does not mean free operation.
Published models cannot be recalled, and built-in safety limits can be removed with fine-tuning, so they can also be misused.
Licenses vary greatly: one may allow almost everything, while another prohibits commercial use or specific application areas.

Similar, but not the same

open-source AI – Open-source AI makes the training code and detailed descriptions of the training data available beyond the weights, and permits free use.
closed model – The weights of a closed model stay with its developer; you can only use it through their service or API.
free AI service – You can use the free chatbot at no cost, but the model runs on the provider's servers and cannot be downloaded.

Related concepts

Broader

AI-modell, foundation model, open AI

Narrower

open-source AI

On the edge

helyben futtatott modell, small language model, closed model, model weights, fine-tuning, quantization

A brief history

The term spread in the first half of the 2020s, as the weights of a growing number of large language models became downloadable while training data remained closed. The Open Source Initiative released its definition of open-source AI in October 2024, clarifying that publishing weights alone does not constitute open source.

Where things stand

At the end of 2026, most of the major AI developers also offer an open-weight model, and in many tasks these come close to the performance of closed models. However, most of the strongest top-tier models remain closed.

Status: 2026 November

In the news

Appears in 19 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Revolut switches from Claude Code to Chinese open-weight model, cutting costs by 90%Nov 3, 2026 · Code
DeepSeek's agent platform reaches 1.54 million daily users as Chinese open models spreadNov 2, 2026 · Agent Age
Mistral Large 4 is the best open model outside China, with weights released publiclyOct 29, 2026 · Agent Age
Mistral's open-weight Le Chonk model ranked alongside Codex GPT-6.1 Terra in codingOct 29, 2026 · Code
European financial firms move to open AI models run on their own infrastructureOct 22, 2026 · Money
Mistral unveils 1-trillion-parameter Large 4, plans open-weights releaseOct 6, 2026 · Agent Age
Together Link runs open models inside Claude Code; Together AI says coding costs could fall by over 50%Oct 5, 2026 · Code
OpenAI DevDay: Codex gets reusable cloud environments that work across devicesSep 29, 2026 · Code

Also known as

open-weight model · open-weight AI · open-weight AI model · open-weights model · open-weight model · open weights · open-weight AI · open model

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.