Llama 4 is born – Long live Llama 3 in the EU

Meta is blocking the multimodal Llama 4 in the EU. Innovation in AI is arriving – but not for us. Entities within the EU are out of luck.

^{Czech version of this post too (Česká verze článku je zde.)}

^{Wikisources – Lama glama Laguna Colorada – CC BY-SA 4.0, Eduardo Avaroa, Bolivia.}

It’s been a year since the release of the Llama 3 model by the American company Meta, which primarily operates Facebook and also Instagram. Simply put: in terms of quality, Llama 3 was for open source in spring last year roughly what the commercial ChatGPT 4 had been a year earlier.

It’s a model with open weights that, at least legally, almost anyone can download and run in their own computing environment. This is important not only for entities that won’t entrust their data to any third party, but it also has a significant impact on markets and competition.

Llama 3 is also served by service providers without them needing to pay license fees or other development costs to the producer. This, in turn, puts downward pressure on the price of commercial models. Moreover, Llama 3 offered an opportunity for Czech startups and the academic world to fully use a fairly capable language model – including very decent Czech – having full access to it both technically and legally (there were legal restrictions even last year, but they were tolerable).

Llama 3 also made it easy to carry out follow-up academic research. For example, the Allen Institute transformed it through post-training into a model called Tülu3, capable of matching or outperforming commercial models.

In January 2025, the Chinese model DeepSeek R1 appeared, which is generally claimed to be more powerful. Whether it actually is will be partially debatable, as the class of thinking models it belongs to is quite specific. For many questions, these reflective models are overkill – like using a cannon to shoot a sparrow – and the more accurate reasoning comes at the cost of slower output. The creators of Tülu3 claim it even outperforms DeepSeek R1 in benchmarks. All three of these models, however, are truly open source by nature.

Commercial models typically offer various versions to choose from, with the thinking versions more of a bonus than intended for main usage.

Llama 4

Last Saturday, Meta rolled out the new Llama 4.

                                      LMarena Leaderboard

This caused even significant changes in the most prestigious chatbot arena ranking (above), where Meta has moved up to second place*. Google is in the lead, and a new version of ChatGPT is expected soon, which will likely shift the rankings once again.

A year ago, I would have told you that the LMarena ranking is entirely reliable and objective. Today, I’m less sure.

The main reason is the fragmentation of testing. The basic knowledge tests are more or less exhausted. Testing is now moving into niches that may be individually valuable but are unnecessary for most everyday users. The same applies to reasoning models. This means the ranking is still valuable at first glance, but ultimately, I’d recommend that everyone try the features themselves and choose the service that best fits their field or domain.

Llama 4 brings many technical innovations. The overarching model, Behemoth (2TB), is almost an order of magnitude larger than Meta’s previous largest version. The context length has increased from last year’s 8K (Llama 3) up to 10M (Llama 4 Scout), although there are doubts about how practical that length really is. It uses an internal mixture-of-experts (MoE) architecture. Models in the new family are “natively multi-modal.“

That last point means they can take text, images, or a combination of both as input.

Llama 4’s license has remained open source in nature. Article 1(a) of the license agreement [2] grants permission “to use, reproduce, distribute, copy, create derivative works of, and make modifications“.

With the exception of subjects from the European Union!

An integral part of the licensing agreement is the Acceptable Use Policy for Llama 4 [3], which states: “With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union. This restriction does not apply to end users of a product or service that incorporates any such multimodal models.“ (emphasis added by the author)

However, the entire Llama 4 model family released on Saturday was presented as natively multimodal. And all ten versions released so far—Scout and Maverick—list their inputs as Multilingual text and image, meaning two distinct modalities. They are thus multimodal and fall under the exclusion cited above.

It is hypothetically possible that a special EU version of the models could be released later, which would be monomodal, for example, by excluding certain internal experts. Given the circumstances outlined below, however, I do not consider this either likely or attractive.

Why?

The online community, particularly in Europe, has understandably responded with complaints about missed opportunities for startups, researchers, and academia. The general belief is that the reason lies in concerns over the recent EU AI Act [6].

In my legal opinion, that’s probably not the main cause.

Meta had already indicated a mismatch with the EU regulatory environment last summer [5], even before the aforementioned act came into effect.

The issue was said to involve the stance of privacy protection regulators—apparently various national data protection authorities, likely acting under the GDPR.

The dispute concerned the usability of user-generated content on Facebook and Instagram (such as posts and photos) as training data for future models. In May last year, Meta contacted all its Facebook users [4], offering them the opportunity to object to the use of their public posts and images for AI purposes training (see below).

Either the European authorities considered this approach unusable (an explicit opt-out instead of explicit consent), or they may have asserted that the right to privacy cannot be waived at all.

Meta, in any case, had already imposed similar licensing restrictions on multimodal models last year with the release of Llama 3.2 and Llama 3.3—using clauses nearly identical to the one mentioned above.

There is some luck in misfortune, for instance with the model Llama-3.3-70B-Instruct from December 2024. Since it is a purely text-based model, the clause does not apply to it. In contrast, the Llama-3.2-90B-Vision-Instruct model processes Text + Image, making it multimodal, so the clause does apply.

The clause also did not appear at all in the earlier Llama 3.1 and Llama 3 models, so they remain usable without these issues.

Exceptions

1. End user of a product containing Llama
2. Local employee of a company operating outside the EU

The first exception is stated directly after the exclusion clause (see above). It doesn’t matter whether the end user uses the service privately or professionally (according to FAQ [7]). However, the service must be more complex—merely hosting the model and indirectly accessing it remotely is not enough.

The second exception is described by Meta on its FAQ page [7]. According to their interpretation, the appropriate license is held by a company operating outside the EU, which can extend it to its employee for the purpose of the company’s activities, which must also be based outside the EU. The FAQ explicitly states that this includes programmers and other developers.

No other exceptions are mentioned. There is no exemption for research, non-profit, educational, or similar use.

What’s next?

A reminder: Every Facebook user has had, and still has, the option to object to the processing of texts and photos they’ve posted for AI purposes, at this address:
https://www.facebook.com/help/contact/6359191084165019

Unlike the various chain hoaxes declaring a ban on user’s data use this kind of objection would actually be taken into account. In other words, user sentiment regarding data use for model training will continue to be reflected.

It’s not entirely impossible that the legal views of the stakeholders may evolve over time. Experience with the new models will become available, including assessments of how much they actually affect privacy. However, such changes won’t happen quickly.

––

You can share your views on the topic in the discussion.

To get to me by email, use the contact info on ILAW.

––

[0] The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation, Meta, Apr 5th 2025 ai.meta.com link

[1] Tülu3, Ai2. allenai.org link

[2] LLAMA 4 COMMUNITY LICENSE AGREEMENT, Meta github link

[3] Llama 4 Acceptable Use Policy, llama.com link

[4] Bringing Generative AI Experiences to People in Europe, May 22, 2024. fb.com link

[5] Meta Withholds Models From Europe, Jul 24, 2024. deeplearning.ai link

[6] REGULATION (EU) 2024/1689 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) eur-lex.europa.eu link

[7] Troubleshooting & FAQ - Restriction on Llama Multimodal Models in the EU llama.com/faq

* Added April 9: In the time since publication, it has come to light that Meta used the model Llama-4-Maverick-03-26-Experimental for its LMarena ranking evaluation—a model different from the ones that were actually released. However, many other developers follow a similar approach, submitting various experimental versions to the ranking tests, which are only available to a small subset of their users.

From one perspective, it is precisely these experimental versions that drive various forms of progress, and comparing them does make sense. At the same time, experiments may suffer from unintended bugs that make them unsuitable for real-world deployment, so it’s legitimate not to pollute production with them and to only gradually drift experiments into deployed systems. From another angle, users are mostly interested in the exact version they actually use—and view such rankings as a bit of a trick.

The model release process has also split into two phases. First, the model’s existence is announced, including its claimed capabilities, and only several months later is it actually deployed for use. This is likely intended to stabilize the user base and discourage too much switching.

This is also why I’ve already recommended taking the ranking with a grain of salt and focusing on your actual needs and go with the solution that fits them.