The landscape of open-weights and open-source language models has shifted significantly in recent years. Anyone looking to understand broader market movements can explore the market developments in open versus closed models to study in depth the commercial and strategic background driving these dynamics. Since the launch of Meta's Llama 3 in spring 2024, fundamental changes have taken place across legal frameworks, applied license models, and community expectations surrounding openness in artificial intelligence.
Those who recall what the market looked like in 2024 know that Llama 3 stood as the open-weights benchmark back then; this post updates that picture to show how the industry has since shifted toward more permissive licenses and stricter open-source definitions. License terms change continuously; always verify the model card or official license text for the current status (verified on 2026-08-07). For developers looking to broaden their perspective on wider trends, the page covering key developments in open-source LLMs is also a valuable resource for the latest insights into adoption and community growth.
Additionally, it is crucial to recognize how technological infrastructure goes hand in hand with legal openness. To understand how modern dataset foundations align with legal frameworks, reviewing the background on open data initiatives for AI provides valuable insight into how open data serves as the basis for compliant training.
Compared to the landscape around the Llama 3 launch, four fundamental shifts in licensing and legal definitions have reshaped the broader industry:
When Meta released Llama 3 on April 18, 2024, it did so under the "Meta Llama 3 Community License". While this license was free to use for the vast majority of developers, it contained strict limitations. For instance, the model could not be used by organizations with more than 700 million monthly active users without a special, separate commercial license from Meta. Additionally, it was mandatory to display the attribution "Built with Meta Llama 3" and the model name had to begin with "Llama 3".
A crucial part of that agreement was the embedded Acceptable Use Policy, designed to prevent misuse. The most debated point at the time, however, was the prohibition against using the weights or generated output to train or improve other large language models. According to the Llama 3 license text this was legally watertight to prevent competitors from directly benefiting from Meta's R&D efforts, which sparked heated debate in the tech world at the time about the true meaning of openness.
Despite its widespread use within the community, the Open Source Initiative (OSI) refused to recognize Meta's licenses as true open source. In their publication, they made it crystal clear that the license suffered from fundamental shortcomings. According to the OSI blog of February 18, 2025 the Llama license fails on three counts: Freedom 0 (the lack of unconditionally free use), OSD point 5 (discrimination against certain users or applications), and OSD point 6 (restrictions on the fields of endeavor for the technology).
To provide a formal foundation for this high-profile discussion, the OSI published the Open Source AI Definition (OSAID) v1.0. Since then, this definition has served as the definitive international benchmark. A model may only call itself open source if it provides, among other things, detailed information about the training data used — although releasing the raw training dataset itself is not always mandatory due to privacy and copyright concerns.
The introduction of OSAID v1.0 brought a sharp demarcation to the market. Where marketers previously dared to label almost any model with publicly accessible weights as "open source," OSAID demands a holistic approach to openness. It is not merely about the binary weights (the parameters), but about a complete ecosystem of transparency. This raises the question of why so many popular open-weights models fail to comply in practice.
The primary stumbling blocks of OSAID v1.0 for many commercial open-weights models lie in data transparency. The standard requires the provider to offer a detailed description of the training data. This encompasses not only the size and composition of the dataset, but also the filtering techniques used, the data sources, and any steps taken to protect copyrighted material or personal privacy. Many major players refuse to disclose this granular information out of fear of legal claims from copyright holders or to safeguard their competitive edge against rivals. Consequently, from a legal standpoint, many models fall back into the "open-weights" category rather than true "open source."
With the release of Llama 4 on 5 April 2025, Meta adjusted its licensing structure on a crucial point. The new "Llama 4 Community License" retained familiar elements from the past, such as the threshold of 700 million monthly active users, mandatory attribution ("Built with Llama"), and the embedded Acceptable Use Policy governed by the laws of the State of California.
The most striking difference in the Llama 4 license text is, however, that the notorious clause prohibiting the use of the models or their outputs to improve other LLMs is entirely absent. In doing so, Meta responded to criticism from researchers and smaller enterprises looking to experiment with model distillation and hybrid architectures, simplifying experiments with model distillation and derivative models.
While early generations of open models almost exclusively operated under proprietary, restrictive community licenses, the market took a different path in 2025. The launch of DeepSeek-R1 in January 2025 stood out because the model combined state-of-the-art performance with the permissive MIT license, as stated on the corresponding Hugging Face model card. This eliminated numerous legal hurdles for commercial implementation.
Not long after, Alibaba followed with the introduction of Qwen3 in April 2025. The Qwen3-30B-A3B model was released directly under the Apache 2.0 license. These developments ensured that developers looking for flexible building blocks can increasingly turn to overviews such as those found under open-source LLMs and the most prominent models, where legal conditions have been greatly simplified and risks are kept to a minimum.
In Europe as well, the trend from tightly controlled licenses toward open standards continued. When Mistral AI announced the Mistral 3 family on December 2, 2025 — consisting of Mistral Large 3 (a Mixture of Experts model with 41 billion active parameters out of a total of 675 billion) and the compact Ministral 3 variants of 3B, 8B, and 14B — they opted for an open strategy.
As shown in the official Mistral 3 announcement, all of these models were released under the Apache 2.0 license. With this move, the company definitively moved away from the more complex commercial frameworks of its early days, paving the way for large-scale industrial adoption within the European Union and beyond.
Alongside the creators' license texts themselves, regulation plays an increasingly significant role in how open models are deployed within the European market. Article 53(2) of the European AI Act contains a specific and valuable exception for open-source and free models. Providers of General Purpose AI (GPAI) models published under an open-source license whose parameters, weights, architecture, and usage information are publicly accessible are exempt from several heavy administrative obligations.
To understand precisely how this exemption works, it is useful to look at what is and is not covered by the regulation. The Article 53(2) exemption relieves open-source providers from the obligation to draw up detailed technical documentation in accordance with the standard GPAI requirements that apply to proprietary models. However, this exemption is not absolute: as soon as an open model is classified as having systemic risk — based on the statutory threshold of more than 10 to the power of 25 FLOPs of cumulative training compute — part of the exemptions lapse, and stringent reporting and evaluation obligations to the European AI Office still apply. Furthermore, copyright policy compliance under Article 53(1) remains fully in force for all models, regardless of their open license.
For a Dutch developer, startup, or enterprise seeking to deploy an open model or fine-tune it for specific use cases, this shift carries far-reaching practical consequences. Simply downloading weights from somewhere and hoping for the best is legally risky. Anyone looking to build seriously must use a structured checklist when analyzing a model card.
When assessing a model card, pay attention to the following core aspects:
Verifying the legal status of an AI model requires a methodical approach. Because repositories on platforms such as Hugging Face or GitHub can change rapidly, it is wise never to rely blindly on third-party summaries. Always start by directly inspecting the file named LICENSE or LICENSE.txt in the root of the model repository. Check whether the stated license is identical to the officially recognized texts of the Open Source Initiative or the Free Software Foundation.
In addition, it is advisable to review the accompanying documentation and the legal/compliance section in the model card. Pay close attention to any addenda or custom terms the creator may have added. If you are uncertain about the implications of a specific clause for your commercial application in the Netherlands, consult a legal professional specialized in open-source compliance and intellectual property. By adopting a proactive and thorough approach, you avoid legal claims down the road and build on a reliable foundation.
The shift from restrictive community licenses toward established open-source standards like MIT and Apache 2.0 marks a maturation of the AI industry. Whereas companies in the early era of generative AI attempted to cover every legal angle with bespoke terms, market forces and regulatory requirements are now demanding clarity. Anyone following the broader trajectory of open technologies will see that the focus is moving from purely owning model weights to data transparency and compliant documentation.
For organizations looking to capitalize on this dynamic, it is crucial not to rely blindly on outdated assumptions. Checking license terms with every new model release remains essential, as legal departments at tech giants continue to refine their conditions. However, those who opt for models licensed under Apache 2.0 or MIT build on a sturdier and far more predictable foundation than was the case in the early days of Llama 3.
The key developments in open-source LLMsDiscover the latest trends in open-source Large Language Models: from efficient small models and…
Quantization in production: from 16-bit to 4-bit precisionPractical guide to LLM quantization in production. Discover the impact of FP16 to INT4 on VRAM, latency…
The race for context: why models are getting increasinglyDiscover why AI models are getting increasingly larger context windows, the impact on web applications…