Guide
What Is Open-Source AI?
A plain-English guide to open-source AI: what it means, how open weights differ from open source and from closed models, how model licences compare, and why it matters for cost, privacy, control and safety.
Definition: Open-source AI is an AI system released so that anyone can use, study, modify and share it, with the model’s weights, the code to train and run it, and detailed information about its training data all published. Most models called “open source” release only the weights, under licences with conditions. Those are open-weight models.
The difference decides what you can do with a model. An open-weight model can be downloaded, run on your own hardware and fine-tuned, but it cannot be rebuilt from scratch, and its licence may limit who can use it commercially. The Open Source Initiative, the body that maintains the best-known definition of open-source software, published its Open Source AI Definition 1.0 on 28 October 2024, and TechCrunch reported that few models described as open source meet it.
This page explains the terms. It does not rank models. For the strongest models you can download, see best open-source LLMs. For what runs on your own computer, see best local LLMs.
What is open-source AI?
An AI model is more than one file. Four parts decide how open it is:
| Part | What it is | Why it matters |
|---|---|---|
| Weights | The billions of numbers the model learned during training, published as a file you can download | Without them you can only use the model through the company’s app or API |
| Code | The software used to train the model and to run it | Lets others check how the model was built and repeat the work |
| Training data | The text, images or other material the model learned from | Decides what the model knows and what biases it carries; needed to rebuild it |
| Licence | The legal terms attached to the release | Decides who may use, change and sell the model, and on what conditions |
A fully open-source model publishes all four, under a licence that lets anyone use it for any purpose. An open-weight model publishes the weights and a licence, and usually the code to run it, but not the training data or the full training code. A closed model publishes none of them: you reach it only through the company’s own products.
The Open Source Initiative definition
The Open Source Initiative (OSI) announced version 1.0 of its Open Source AI Definition at the All Things Open conference in Raleigh, North Carolina, on 28 October 2024. It says an open-source AI system must grant four freedoms:
- Use the system “for any purpose and without having to ask for permission”.
- Study how the system works and inspect its components.
- Modify the system for any purpose, “including to change its output”.
- Share the system for others to use, with or without changes, for any purpose.
To make those freedoms practical, the definition requires three things to be released:
| Requirement | What the OSI definition asks for |
|---|---|
| Data information | ”Sufficiently detailed information about the data used to train the system” for a skilled person to build a substantially equivalent system |
| Code | ”The complete source code used to train and run the system” |
| Parameters | ”The model parameters, such as weights or other configuration settings” |
The definition does not require the training data itself to be published, only enough information about it to rebuild a similar system. It states that “Open Source models and Open Source weights must include the data information and code used to derive those parameters.” On that reading, weights alone are not open source.
Not every company accepts the definition. Meta has called its Llama models open source, as in Mark Zuckerberg’s July 2024 letter “Open Source AI Is the Path Forward”. When the OSI definition was published, a Meta spokesperson told TechCrunch: “we, like others across the industry, disagree with their new definition.”
Open source vs open weights vs closed AI
| Fully open source | Open weights | Closed | |
|---|---|---|---|
| Weights downloadable | Yes | Yes | No |
| Training code published | Yes | Sometimes, partly | No |
| Training data or detailed data information published | Yes | Rarely | No |
| Run on your own hardware | Yes | Yes | No |
| Fine-tune on your own data | Yes | Yes | Only through the company’s own tools, where offered |
| Rebuild the model from scratch | Possible, with enough computing power | No | No |
| Licence | Permits any use | Ranges from permissive to restricted | The company’s terms of service |
| Example | Olmo 3 from Ai2 | Llama 4, DeepSeek V4, gpt-oss | The models behind ChatGPT, Claude and Gemini |
Ai2, the Allen Institute for AI, describes its Olmo models as “our fully open language model and complete model flow”, and publishes the weights, the training and post-training code, and the pre-training, mid-training and post-training datasets.
What “open weights” means
Open weights means the trained model file is published for anyone to download. A weights file is the finished product of training: the numbers a neural network uses to turn an input into an output. With it, you can run the model on your own computer or server, fine-tune it on your own data, and build it into your own products, within the licence’s terms.
What you cannot do is see exactly how it was made. Without the training data and full training code, nobody outside the company can check what the model learned from or train the same model again. DeepSeek’s model card for DeepSeek-V4-Pro is typical: it publishes the weights under the MIT licence and says the model was trained on “more than 32T diverse and high-quality tokens”, but it does not release that data.
Companies use different words for the same kind of release. Google introduced Gemma in February 2024 as “open models”, not open source. OpenAI called gpt-oss, released on 5 August 2025, its “first open-weight language models since GPT‑2”.
Open-source AI licences compared
The licence matters as much as the weights. Some are standard open-source software licences with no extra conditions. Others are written by the model’s maker and add limits on large companies, naming or attribution.
| Model | Maker | Licence | Main conditions | Source |
|---|---|---|---|---|
| gpt-oss | OpenAI | Apache 2.0 | Standard Apache 2.0 terms: keep the copyright and licence notice | OpenAI |
| Gemma 4 | Apache 2.0 | Standard Apache 2.0 terms: keep the copyright and licence notice | ||
| Qwen 3.6 | Alibaba | Apache 2.0 | Standard Apache 2.0 terms: keep the copyright and licence notice | Hugging Face |
| DeepSeek V4 | DeepSeek | MIT | Standard MIT terms: keep the copyright and licence notice | Hugging Face |
| Mistral Large 4 | Mistral | Mistral Ecosystem Open-Weight License, which Mistral calls a modified MIT licence | A business offering the model as a hosted service needs a separate commercial licence from Mistral | Mistral |
| Kimi K3.1 Pro | Moonshot | Kimi K3.1 Pro License, which grants MIT-style rights with added conditions | A separate agreement for hosted-model businesses with more than $20 million revenue over 12 months; products with more than 100 million monthly users or $20 million monthly revenue must show the model’s name | Hugging Face |
| Llama 4 | Meta | Llama 4 Community License | Companies with more than 700 million monthly active users must ask Meta for a licence; products must show “Built with Llama”; models trained from it must start their name with “Llama” | Meta |
| Muse Spark Open | Meta | Muse Model License | What Meta calls “limited commercial terms, which allow many developers to use the model for free” | Meta |
Apache 2.0 and MIT are long-standing open-source software licences that let anyone use, change and sell the work, as long as the copyright notice is kept. Custom licences vary. Before building a product on a model, read its licence in full, including any acceptable-use policy it points to.
None of the models in this table publishes its full training data, so none of them meets the OSI definition, however permissive its licence.
Why open weights matter: cost, privacy and control
Open weights change who controls a model once it is released.
- Privacy and data location. A model running on your own hardware never sends your prompts or documents to the company that made it. That matters for organisations bound by data-protection rules or client confidentiality.
- Cost. There is no per-token bill, but you pay for the hardware and the electricity. OpenAI says gpt-oss-120b runs on a single 80 GB GPU and gpt-oss-20b on devices with 16 GB of memory. For the cost of renting the same models from a hosting company, see best LLM APIs.
- Control. A downloaded model cannot be changed, withdrawn or retired by its maker. Your copy keeps working as it is.
- Customisation. You can fine-tune the weights on your own data to suit a specialist task.
- Offline use. Smaller open-weight models run on a laptop with no internet connection. The NPU guide explains the chips that run AI on a device.
Meta’s chief executive, Mark Zuckerberg, set out the case for open models in a July 2024 letter: “We need to control our own destiny and not get locked into a closed vendor.” He said developers could run Llama 3.1 405B on their own infrastructure at about half the cost of using closed models such as GPT-4o. That is Meta’s own estimate.
The trade-off is work. Someone has to choose the hardware, install the model, keep it running and apply updates. Using a closed model through an app or API leaves all of that to the company.
The safety debate around open weights
Whether powerful models should be released as open weights is disputed.
The International AI Safety Report, the work of 96 international AI experts, published in January 2025, sets out both sides. It says open-weight access “facilitates research and innovation, including in AI safety, as well as increasing transparency and making it easier for the research community to detect flaws in models”. It also says open-weight models “can also pose risks, for example by facilitating malicious or misguided use that is difficult or impossible for the developer of the model to monitor or mitigate”, and that “once model weights are available for public download, there is no way to implement a wholesale rollback of all existing copies”. The report says experts have broadly agreed to judge openness by its marginal risk: how much releasing a model adds to the risks that closed models and other technologies already pose.
| Position | Who has argued it | Source |
|---|---|---|
| Open models are safer because more people can inspect them | Meta’s Mark Zuckerberg: “open source should be significantly safer since the systems are more transparent and can be widely scrutinized” | Meta, July 2024 |
| Governments should monitor open-weight models rather than restrict them | The US National Telecommunications and Information Administration recommended that the government “refrain from immediately restricting the wide availability of open model weights in the largest AI systems” while it builds ways to monitor risks | NTIA, 30 July 2024 |
| Once released, weights cannot be recalled if a problem is found | The International AI Safety Report: “there is no practical way to roll back access if it is later discovered that a model has faults or capabilities that enable malicious use” | International AI Safety Report, January 2025 |
Open-source AI and the law
The European Union’s AI Act gives some open-source models lighter rules. Article 53 exempts providers of general-purpose AI models from two duties, keeping technical documentation and supplying information to companies that build on the model, when the model is “released under a free and open-source licence that allows for the access, usage, modification, and distribution” of it and its weights, architecture and usage information are public.
The exemption has a limit: “This exception shall not apply to general-purpose AI models with systemic risks.” The most powerful open models carry the same obligations as closed ones.
How to use an open-weight model
There are three common ways to use one:
| Route | What it involves | Where to start |
|---|---|---|
| Through a hosted API | A cloud company runs the model and charges per token | Best LLM APIs |
| On your own computer | A desktop app downloads and runs a model sized to your memory | Best local LLMs |
| On your own servers | Your team hosts the model on its own GPUs for full control over data | Best open-source LLMs |
For how a model turns text into tokens and what that costs, see AI tokens. For how models are measured against each other, see AI benchmarks.
Frequently asked questions
What is open-source AI?
Open-source AI is an AI system released so that anyone can use, study, modify and share it for any purpose. Under the Open Source Initiative’s Open Source AI Definition 1.0, published on 28 October 2024, that means releasing the model’s weights, the complete code used to train and run it, and enough information about its training data for a skilled person to build a similar system.
What is the difference between open source and open weights?
Open weights means the trained model file can be downloaded and run. Open source, under the Open Source Initiative’s definition, also requires the training code and detailed information about the training data, so the model can be studied and rebuilt. Most models described as open source, including Llama, DeepSeek and Qwen, are open-weight models.
Is Llama open source?
Meta has called Llama open source, but Llama 4 is released as open weights under the Llama 4 Community License, which has conditions. Companies with more than 700 million monthly active users must ask Meta for a licence, and products must show “Built with Llama”. Meta told TechCrunch in October 2024 that it disagrees with the Open Source Initiative’s definition.
Is ChatGPT open source?
No. ChatGPT is an app from OpenAI, and the models behind it are closed: their weights are not published. OpenAI has released separate open-weight models, gpt-oss-120b and gpt-oss-20b, under the Apache 2.0 licence on 5 August 2025, which it called its first open-weight language models since GPT-2.
Which AI models are fully open source?
Very few. Ai2’s Olmo models are released with their weights, training code and training datasets, and Ai2 describes them as fully open. Popular open-weight models such as DeepSeek, Qwen, Gemma, gpt-oss and Llama do not publish their training data, so they do not meet the Open Source Initiative’s definition.
Can I use open-source AI models commercially?
Usually, but it depends on the licence. Models under Apache 2.0 or MIT, such as gpt-oss, Gemma 4, Qwen 3.6 and DeepSeek V4, can be used commercially under the standard terms of those licences, which ask you to keep the copyright and licence notice. Custom licences, such as Meta’s for Llama or Mistral’s for Mistral Large 4, add conditions for very large companies or for businesses that host the model for others.
Is open-source AI safe?
It is disputed. The International AI Safety Report (January 2025) says open weights make it easier for researchers to find flaws, but that misuse is hard for the developer to monitor and that released weights cannot be recalled. Meta argues open models are safer because more people can inspect them. The US NTIA recommended monitoring rather than restricting them in July 2024.
Why would anyone use an open-weight model instead of ChatGPT or Claude?
The main reasons are privacy, control and cost. An open-weight model running on your own hardware does not send data to its maker, cannot be withdrawn or changed by the company, can be fine-tuned on your own data and has no per-token bill. The trade-off is that you must provide the hardware and keep the model running yourself.