TL;DR
– “Open source AI” has a formal definition. The Open Source Initiative released the Open Source AI Definition 1.0 on 28 October 2024. It asks for usable training data information, full code and the model parameters.
– Open weights is not the same thing. Many popular models publish their weights but keep training data and code private, or add licence restrictions.
– The foundations matter more than any yearly ranking. Frameworks, model hubs, open-weight model families and local tools each play a different role.
– Licences differ a lot. Some models use Apache 2.0 or MIT. Others use custom licences with conditions, so read the licence before you build on a model.
– EU law treats open models a bit differently. The AI Act lifts some documentation duties for openly licensed general-purpose models, but not for models with systemic risk.
Contents
1. What open source AI actually means
2. Open weights vs open source
3. Frameworks and libraries
4. Open-weight model families
5. Tools for running AI locally
6. Why it matters and how to judge a project
7. FAQ
When this article first appeared, “open source AI” mostly meant research code and corporate labs sharing tools. Since then, open models have become something people download and run on a laptop, and the term has turned into an argument. This guide skips the “top 5 of the year” format, which ages badly. Instead, it explains what the label means, groups the projects that have shaped the field by the role they play, and gives you criteria to judge new ones yourself.
What open source AI actually means
Open source software has had a shared definition for decades: you can use, study, change and share the code. AI systems are harder. A model is not only code. It is also the trained parameters (the “weights”) and the data that produced them.
The Open Source Initiative (OSI), the body that maintains the classic Open Source Definition, published the Open Source AI Definition 1.0 on 28 October 2024, after what it called “a year-long, global, community design process”. It keeps the four familiar freedoms (use, study, modify, share) and spells out what you must receive to exercise them.
| Component | What the OSI definition requires |
|---|---|
| Data information | “Sufficiently detailed information about the data used to train the system so that a skilled person can 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” |
Note what it does not demand: the raw training data itself. Detailed information about the data is enough. That compromise is one reason the definition was debated.
Open weights vs open source
Most well-known “open” models are better described as open weights. You can download the trained model and run it, but you often get little about the training data, sometimes no training code, and sometimes a licence with strings attached.
That is not a scandal. Open weights are still useful: you can run the model privately, fine-tune it and inspect its behaviour. But it is a different promise from open source in the OSI sense, because you cannot fully study or reproduce how the model was made.
| Open source AI (OSI sense) | Open weights | Closed / API-only | |
|---|---|---|---|
| Download the model | Yes | Yes | No |
| Run it on your own hardware | Yes | Yes | No |
| Training code | Yes | Sometimes | No |
| Training data information | Detailed | Usually limited | Usually limited |
| Licence | OSI-style, no use restrictions | Varies, from permissive to restrictive | Terms of service |
When a company calls its model “open”, ask which column it actually belongs in.
Frameworks and libraries
These are the layers most AI work is built on. They are fully open source code, and they tend to outlast any single model.
| Project | Role | Why it has staying power |
|---|---|---|
| PyTorch | Deep learning framework | Started at Meta. Since 12 September 2022 it lives under the PyTorch Foundation at the Linux Foundation, so no single company controls it |
| TensorFlow | Deep learning framework | Google’s framework, with a mature set of tools for deploying models |
| JAX | Numerical computing and machine learning | Research library for high-performance numerical code on GPUs and TPUs; its README says it is not an official Google product |
| Hugging Face Transformers | Model library | A common interface for loading and using pretrained models from many developers |
| Hugging Face Hub | Model and dataset hub | A main place where open-weight models are published, with model cards and licences |
| scikit-learn | Classic machine learning | Still the standard starting point for non-deep-learning tasks in Python |
The original version of this article recommended scikit-learn for beginners. That advice holds.
Open-weight model families
Model releases arrive fast, so names below are families, not specific versions. The licence details are what trip people up, so here they are with sources.
| Family | Developer | Licence nuance (example release) |
|---|---|---|
| Llama | Meta | Custom licence. Under the Llama 3.1 licence, companies whose products had more than 700 million monthly active users must “request a license from Meta, which Meta may grant to you in its sole discretion” |
| Mistral | Mistral AI | Mistral 7B was released under Apache 2.0 and, in the company’s words, “can be used without restrictions” (Mistral AI, 2023). Check each later model separately |
| Qwen | Alibaba Cloud | For Qwen2.5, “all our open-source models, except for the 3B and 72B variants, are licensed under Apache 2.0” (Qwen team) |
| DeepSeek | DeepSeek | DeepSeek-R1 code and models were released under the MIT licence, with distillation and commercial use allowed (DeepSeek, January 2025) |
Two practical points. First, licences change between versions of the same family, so check the model card for the exact release you use. Second, a permissive licence on the weights still does not make a model open source in the OSI sense if the training data information is thin.
Tools for running AI locally
This layer made open models practical for people without a data centre.
- llama.cpp runs language models efficiently on ordinary CPUs and consumer GPUs. Many other tools build on it.
- Ollama wraps local models in a simple download-and-run experience, with a local API other apps can call.
- Stable Diffusion brought open image generation to home computers. Check the licence of the exact version you use.
Local use matters for privacy (data does not leave your machine), for cost (no per-request fees once you have the hardware) and for resilience (your workflow does not depend on one provider’s API). It also matters as AI agents become more autonomous: an agent built on a model you control is easier to audit. For more on that direction, see our explainer on the synthetic economy.
Why it matters and how to judge a project
Transparency. Open weights let researchers test models for bias, safety issues and hidden behaviour. Open data information goes further and lets them check what the model learned from.
Cost and control. You can fine-tune a model on your own data and run it on your own hardware, instead of renting access.
Regulation. The EU AI Act gives openly licensed general-purpose models a partial break. Under Article 53(2), the duties to keep technical documentation for authorities and to supply documentation to downstream providers do not apply to models released under a free and open-source licence whose weights, architecture and usage information are public. Two catches: providers must still have a copyright policy and publish a summary of training content, and “this exception shall not apply to general-purpose AI models with systemic risks”. For the AI Act’s labelling rules, see our guide to Article 50.
To judge any project, new or old, ask these questions:
| Question | What a good answer looks like |
|---|---|
| What exactly is released? | Code, weights and training data information, not just a demo |
| What is the licence? | A standard licence (Apache 2.0, MIT) or a custom one you have actually read |
| Who governs it? | A foundation or a broad contributor base, not one company that can change course alone |
| Is it maintained? | Recent releases, answered issues, documentation |
| Can you run it? | Hardware needs that match what you have |
Projects that pass these tests tend to stay relevant. Projects that only pass the press release do not.
FAQ
What is the difference between open source AI and open weights?
Open weights means you can download and run the trained model. Open source AI, as defined by the OSI, also requires the complete training and running code and detailed information about the training data, under terms that let you use, study, modify and share it.
Is Llama open source?
Not in the OSI sense. Meta publishes the weights under a custom licence that, for Llama 3.1, requires very large companies to get separate permission from Meta, while the OSI definition requires freedom to use a system “without having to ask for permission”.
Which open source AI framework should a beginner learn?
PyTorch is a widely used starting point for deep learning, and scikit-learn for classic machine learning in Python. Both are free and well documented.
Can I use open-weight models commercially?
Often yes, but it depends on the licence. Apache 2.0 and MIT models allow commercial use; custom licences may add conditions, so read the licence for the specific version.
Does the EU AI Act apply to open source AI models?
Yes, partly. Openly licensed general-purpose models are exempt from some documentation duties, but not from copyright and training-summary duties, and not at all if the model carries systemic risk.
Sources: Open Source Initiative – Open Source AI Definition 1.0 · OSI – release announcement · EU AI Act, Article 53 · Linux Foundation – PyTorch Foundation · Meta – Llama 3.1 licence · Mistral AI – Mistral 7B · Qwen team – Qwen2.5 · DeepSeek – R1 release
About the Author
👉Want guest posting?🪙Subscribe to the @egorithms Telegram