open-source model
An AI model whose weights, and ideally its code and training details, are made freely downloadable by its creator, allowing anyone to run it on their own machine, examine it, and adapt it for their own purposes within the license framework.
In plain terms
Most well-known AI services are like a restaurant: you order the food, you get it, but you cannot see into the kitchen, and you cannot take the recipe home. Most downloadable models are like taking the prepared meal home and seasoning it further at home. With a truly open-source model, you also get the recipe and the list of ingredients.
In practice, this means the developer publishes the model's "brain", i.e., the mass of numbers (weights) formed during training. You can download it and run it on your own computer or your company's server. Running it no longer requires an internet connection or a subscription.
This is mainly an advantage if you do not want your data to leave your premises, if you do not want to depend on a single provider, or if you want to further train the model for a narrow task. The cost is that operation, hardware, and expertise are your responsibility.
The word 'open' covers many things. In some cases, only the weights are released under strict usage conditions; in others, the code and training details are also included. Therefore, always check the license, not the label.
How it works
- A company or research group trains a model at great cost, then uploads the finished weights to a public repository, such as Hugging Face.
- A license is attached to the model, specifying how you can use it: whether it is allowed for business purposes, if it can be modified, or shared.
- You download the model and run it on your own machine or a rented server using a launcher. Smaller models run on a powerful laptop, while larger ones require serious graphics cards.
- If needed, you fine-tune it with your own data, meaning you continue training it to better understand legal texts or your company's products.
- Meanwhile, the community creates smaller, faster, specialized versions of it and shares them, so an entire family grows from one model.
An everyday example
A law firm would not want to upload clients' confidential documents to an external AI service provider. Therefore, it downloads an open model and runs it on its own server. The summarization of the documents thus remains in-house, and the model works even without internet.
Who is affected
A common misconception
Limits and risks
- The downloadable model can make mistakes and hallucinate facts just like its closed counterparts.
- Running large models requires expensive hardware and expertise, so for small-scale usage, a subscription or a paid API is often cheaper.
- Once a model is released, its creator can no longer recall or supervise it, so others can remove its built-in safety guardrails.
- You are responsible for the update, security, and lawful use, and there is usually no customer support to turn to unless you purchase separate support.
Similar, but not the same
- open-weight model – Here, only the final weights can be downloaded; the training code and data descriptions are not included. Therefore, in the strict sense, this is less than open source, although common usage often conflates the two.
- closed model – A closed model is accessible only through the developer's servers, via an application or API. You cannot download or modify it.
- free AI service – A completely closed model can stand behind a free-to-use chat program. Free refers to the price, openness to access.
Related concepts
language model, foundation model, open-source software
–
open-weight model, closed model, small language model, fine-tuning, local execution, model weights, licenc
A brief history
The term 'open source' was coined in 1998 in the software world and was later transferred to artificial intelligence. It became widespread among large language models from 2023, when several major developers made their models downloadable, with Meta being the first with the LLaMA family. The Open Source Initiative released version 1.0 of the definition of open-source AI in October 2024. This is the first widely referenced definition, but it remains debated to this day.
Where things stand
At the end of 2026, several major American, European and Chinese developers release downloadable models, and these come close to closed models on many everyday tasks. The debate over what counts as truly open is still unresolved.
In the news
- 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
- 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
- Meta launches open-source AI hardware project for makers under Apache 2.0 licenceOct 3, 2026 · Code
- OpenAI DevDay: Codex gets reusable cloud environments that work across devicesSep 29, 2026 · Code
- Open-source AI security agent used to find 24 Android vulnerabilitiesSep 28, 2026 · Code
Also known as
open-source model · open model · open-source modell · open source model · open model · open-source AI
Sources
- Open Source Initiative – The Open Source AI Definition
- Wikipedia – Open-source artificial intelligence
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.