Regulatory friction in Europe
Meta’s refusal to release multimodal Llama models in the EU creates a severe gap for European developers. Meta spokesperson Kate McLaughlin told The Verge that the company will not release multimodal Llama models in the EU because of the "unpredictable nature" of the European regulatory environment, which limits the available tools for local developers. This decision follows Meta’s July 20, 2025, refusal to sign the EU’s voluntary code of practice for general-purpose AI. Joel Kaplan, Meta’s Chief Global Affairs Officer, said the code creates "legal uncertainties for model developers" and exceeds the scope of the AI Act. While Microsoft signed the code to shape governance before the August 2, 2025, mandatory enforcement, Meta maintains its distance. This regulatory friction limits the Llama 4 ecosystem in Europe. The AI Act’s transparency rules take effect in August 2026, and the European Commission released technical guidelines on July 18, 2025, to assist with implementation. Since the 2026 compliance deadline approaches, Meta’s stance creates a divide between the company and European regulators. This pattern of pushback mirrors Apple’s decision to exclude the EU from its AI Intelligence rollout because of the 2022 Digital Markets Act. Apple told Bloomberg that it feared interoperability requirements could force it to compromise product integrity.
Performance leaps and the monetization struggle
The Llama 4 family attempts to challenge closed models through mixture-of-experts architecture. Llama 4 Maverick, which launched April 5, 2025, uses 128 experts to achieve 400 billion total parameters. It beats OpenAI’s GPT-4o on reasoning, coding, and vision tasks. Llama 4 Scout provides a 10 million token context window to help summarize large codebases. I find the technical leaps impressive, yet the financial reality remains grim. In Q1 2026, Meta’s stock tumbled after the company announced increased AI infrastructure spending, unlike Alphabet’s steady reaction to similar capex guidance. Meta’s revenue still relies heavily on Facebook and Instagram ads. You should remember that Meta must prove its massive capital expenditures lead to real revenue growth beyond the metaverse. Can Meta ever prove its AI spending leads to more than just better ad targeting?
| Model | Active Parameters | Total Parameters | Feature |
|---|---|---|---|
| Llama 4 Scout | 17 billion | 109 billion | 10M token context window |
| Llama 4 Maverick | 17 billion | 400 billion | Beats GPT-4o on reasoning |
| Llama 4 Behemoth | 288 billion | ~2 trillion | STEM-focused performance |
Llama 4 Behemoth, which remains in training, contains 288 billion active parameters across 16 experts for a total approaching two trillion parameters. Internal benchmarks show this model outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused tests. Meta aims to use these models to power its AI assistant in WhatsApp, Messenger, and Instagram Direct. The company also intends to offer a commercial version that allows businesses to customize models using proprietary data. This commercial version will compete directly with Microsoft-backed OpenAI and Google. The release of the new commercial model and customization service is imminent.
Competition and security vulnerabilities
Mistral AI continues to compete with Meta using models like Pixtral Large. Mistral’s 124 billion parameter Pixtral Large matches or surpasses Claude 3.5 Sonnet and Gemini 1.5 Pro in multimodal benchmarks. Mistral also offers enterprise licenses for its models, which competes with Meta’s push for commercial Llama versions. Mistral’s flagship chatbot, Le Chat, now includes web searching and inline citations to increase transparency. It also has a canvas tool to modify content and can analyze large PDFs or images. Mistral distributes its models on Google Cloud and Microsoft Azure. The company’s release of Mistral Large 24.11 targets long-context understanding for task automation. Llama 4 Maverick also has flaws. A security report on Maverick found 3 critical and 5 high severity findings across 39 test categories. The model also scored 0% in both Religious Bias and Political Bias categories. These failures in bias mitigation show that Meta still struggles with model safety. The Maverick model shows vulnerabilities in protection against unauthorized access and data exposure, and it failed tests regarding Pliny Prompt Injections. The evaluation of Llama 4 Maverick against OWASP Top 10 and MITRE ATLAS frameworks revealed six findings.
