Open weights vs open source AI: what is the difference?
Open weights means a model's trained parameters can be downloaded. Open-source AI, as the OSI defines it, also needs code and training-data information. Here is why it matters.
Short answer. Open weights means you can download the trained model. Open-source AI, as the Open Source Initiative (OSI) defines it, asks for much more: the code, detailed information about the training data and the weights, all under terms that let you use, study, modify and share the system. Most popular “open” models are open weights, and a model under a permissive license can still fall short of the open-source definition.
Four levels of openness
| Level | What is public | What is usually missing | Examples in our catalog |
|---|---|---|---|
| Weights, research-only license | The trained weights | Training data and code; commercial use | Qwen-Image-2.1 |
| Weights, custom license with conditions | The trained weights | Training data and code; some uses need permission or a fee | FLUX 3 Image Klein, Kimi K3.1 Pro |
| Weights, permissive license | The trained weights under Apache 2.0, MIT or a similar license | Usually the training data and the full training code | Gemma 4, DeepSeek V4.1 Flash |
| Fully open | Weights, training code and checkpoints, plus the training data or the pipelines to rebuild it | Little | Olmo 3.1, Apertus 1.5, K2 Horizon |
There are steps in between. For example, OpenBMB released many of the datasets behind MiniCPM5-2B, which goes beyond a plain open-weights release.
Where “open source” comes from
In software, open source means the source code is available under a license that lets everyone use, study, change and share it. The OSI keeps the list of approved licenses.
AI does not fit that picture neatly. A trained model is produced from three ingredients: training data, training code and computing power. The weights are the result, a huge table of numbers rather than source code that a person can read and edit. Releasing only the weights is a bit like releasing a compiled program without the code that built it: you can run it and, with effort, adjust it, but you cannot rebuild it from scratch.
What the OSI definition requires
The Open Source Initiative approved version 1.0 of its Open Source AI Definition in October 2024. It grants four freedoms:
- to use the system for any purpose without asking permission;
- to study how it works;
- to modify it;
- to share it, with or without modifications.
To make those freedoms real, it requires the preferred form for making modifications, which has three parts:
- Data information: sufficiently detailed information about the training data, so that a skilled person could build a substantially equivalent system. This does not always mean the data itself.
- Code: the complete code used to train and run the system.
- Parameters: the model weights.
Each part must come under terms that grant those freedoms.
Why a permissive license is not enough
Apache 2.0 and MIT are OSI-approved licenses, and they are the friendliest terms you will meet in AI. But a license only covers what is actually released. A model under Apache 2.0 whose developer publishes no data information and no training code is open weights under a permissive license. It is excellent for building products, and it still does not meet the full definition.
The reverse also happens. Terms that limit commercial use by company size or field of use cannot give the four freedoms. The LTX-2.x Community license is a clear case: our license page notes that it is not open source under the OSI definition, because it discriminates by revenue and by field of use. And OpenMDW-1.1, a permissive license written for machine-learning models, has not been approved by the OSI.
Why the difference matters in practice
- What you may do. The license decides whether you can build a business on the model, fine-tune it, redistribute it or run it as a service. Read AI licenses explained before you ship.
- What you can check. Without information about the training data, you cannot check it for copyrighted material, personal data or bias.
- What you can reproduce. Researchers can only rebuild a model, or learn how choices in training change its behavior, when the data, code and checkpoints are public.
- What you can trust in a label. “Open source” on a model page is a claim, not a guarantee. Ask which parts are public.
How this site handles it
On this site, “open model” means open weights, and we say so. Every model page shows the license type: permissive, conditional or non-commercial, with the conditions in plain words. We mark fully open releases, such as Olmo 3.1, Apertus 1.5 and K2 Horizon, with a “Fully open” tag, and we note downloads that are gated on Hugging Face.
If you are new to all this, start with what an open-source AI model is.
Frequently asked questions
- Does an Apache 2.0 license make an AI model open source?
- Not by itself. Apache 2.0 is an OSI-approved license, but the Open Source AI Definition also asks for detailed information about the training data and the code used to train and run the system. A model under Apache 2.0 is open weights under a permissive license, and whether it meets the full definition depends on what else is released.
- Does open weights mean the training data is public?
- Usually not. Most open-weights releases publish the trained model and a license, but not the data it was trained on. A few developers go further: Ai2 publishes the data and code for its Olmo models, the Swiss AI Initiative publishes the data preparation pipelines, training code and checkpoints for Apertus, and IFM at MBZUAI publishes training code, data or data recipes, intermediate checkpoints and logs for K2 Horizon.
- Why do companies call open-weights models open source?
- The term is familiar and carries goodwill from open-source software. Strictly, many of these releases are open weights, which is why this site shows the license type on every page and uses the more precise label.
- Which should I choose, open weights or fully open?
- If you want to run a model and build products, a permissive open-weights model is usually enough. If you need to audit the training data, reproduce the training or do research on how models are made, choose a fully open release.
Key terms in this guide
- Gated model A gated model is an open model whose download on Hugging Face requires you to log in and accept its terms, or request access, first.
- Open-source AI Open-source AI, as defined by the Open Source Initiative, is an AI system you may use, study, modify and share freely, with the weights, the code and detailed information about the training data all available.
- Open weights An open-weights model is an AI model whose trained parameters, called weights, are published so that anyone can download and run it, usually under a license that sets the terms of use.
- Permissive license A permissive license, such as Apache 2.0 or MIT, lets you use, modify and sell a model with very few conditions, usually just keeping a notice.
Models mentioned in this guide
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Olmo 3.1 Allen Institute for AI (Ai2) | 7.3B – 32.2B | 64K | Apache 2.0 (Permissive) | 5.04 GB Laptop or phone | |
| Apertus 1.5 Swiss AI Initiative | 8B – 70B | 256K | Apache 2.0 (Permissive) | 5.52 GB Laptop or phone | |
| K2 Horizon Institute of Foundation Models (MBZUAI) | 1.08B – 379B 23B active | 512K | Apache 2.0 (Permissive) | 0.745 GB Laptop or phone | |
| MiniCPM5-2B OpenBMB | 1.08B – 2.52B | 128K | Apache 2.0 (Permissive) | 0.745 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone | |
| Qwen-Image-2.1 Alibaba Qwen | 7.1B | – | Qwen Research License (Non-commercial) | 20 GB Gaming GPU | |
| Kimi K3.1 Pro Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster |