Meta reveals the Lama 4 AI series with a new expert -based architecture.

Meta, the new AI Model Series on April 5: Llama 4 containing Lama 4 Maverick and Lama 4 Scout, is designed to process speeches and large files with an unpublished “Teacher” model Lama 4 Behemoth respectively.
Lama 4 It is the first series of Meta to adopt a mixture of “MoE) architecture. This approach activates only the elite parts of the nervous network called “experts için to commit certain sub -tasks. The task will be devoted to sub -duties and each will be directed to the most suitable experts and will increase resource efficiency.
What are the features related to Lama 4 Maverick and Scout?
Llama 4 MAVERICK has 128 experts and 17 billion active parameters representing the information of a model used to process a particular entry. Meta describes this as the “Product Labor Model for General Assistant and Chat Examples Specializing in Image Interpretation and Creative Writing.
Interestingly, Mark Zuckerberg’s company boasts that Maverick offers ı cost rate for the best of its class ”when it comes to conversations. Since the surprise version of Deepseek in January, it has been playing in the minds of AI giants, 5.6 million dollars for education.
See: Meta’s offer of $ 800 million to the Chip initiative was rejected – why why
However, the AI experts, the version of Llama 4 Maverick published in LMarena, which listed large language models in various tasks “Optimized for speech” And shows different performance From the public version. This shows that Meta offers a modified version of LMarena that will be higher in the leadership table.
Llama 4 Scout’s 17 billion active parameters and only 16 experts, but Meta says it is the “best multimodal model in the world in class”. It has an unusual context window that represents the amount of information it can process in a request, so it performs well in summarizing large documents and in sequential reasoning.
Meta says that both Scout and Maverick are “best ve and 28 billion active parameters and 16 experts because they are distilled from Lama 4 Behemoth. Already in LMarena, it is at a high level, it is still trained and not released. Accordingly InformationThe Lama 4 announcement was postponed at least twice due to low -performance models in technical criteria and speeches.
How can you access Lama 4 Maverick and Scout?
Scout and Maverick can be downloaded to Llama.com and Huging Face, or can be used in 40 countries via WhatsApp, Messenger and Meta AI Chatbots on Instagram. Multimodal properties can only be used in the USA and in English.
Some partners have already announced their integrations; Developers can create and distribute AI applications with Microsoft’s Llama 4 models in Azure AI Foundry and Azure Databricks.
Lama 4 is apolitical
Meta, Lama 4 models, “removing prejudice” said that it works specially. Rejection of questions about “discussed political and social issues üzerinden over 5% of Lama 3.3 and the questions that decrease are defined as“ significant more balanced ”.
US President Donald Trump’s team expressed skepticism about the neutrality of AI models, AI and Crypto Tsar David Sacks show that Openai is Chatgpt ”Programmed to be awake“In a podcast. AI experts, ultimately prejudice stems from training data and can lead political tendencies In any direction, not just left.
However, Zuckerberg’s company made a series of final moves, which showed that he wanted to stay with the US administration. Republican strategist Joel Kaplan was hired as Meta’s policy leader shortly after Trump took office; He sees the social media arrangement as a direct challenge to free speech.
In January, Meta announced that the company left it. Third Party Real Control Program And to move the content control teams from California to Texas, “helping to eliminate the concern of prejudiced employees that they are overly censored”. After criticizing Trump’s diversity, equality, and attempts to include Trump, he eliminated.
In addition, Meta, Lama 4 models, current White House Advisor Elon Musk’a produced by Xai Chokbot produced by Grook, a “contentious” issues similar to a “strong political simple” issues, he said.
Lama 4 cannot be used in the EU
Accordingly Lama 4 acceptable use policyIn the European Union, individuals with “residing” or companies with “main workplace” cannot use or distribute models. However, these individuals or companies can use Lama 4 models when they are included in a product or service they can access in the region.
This is probably when it comes to AI’s problems with the EU legislation of the Meta. In June 2024, Meta postponed his education. LARGE LANGUAGE MODELS After the EU, the publicly shared on Facebook and Instagram, the regulators argued that the content owners should get a clear approval. Meta AI is still not released in the block.
See: Meta offers less personalized ads for EU users
Meta signed an open letter that called on European regulators to address the demand for ız inconsistent regulatory decisions ”and an unpredictable compliance demands last September. Later, in February, Meta announced that the company was ready to increase its concerns that Trump sees it as unfair EU regulations.
There are other restrictions in the use of Lama 4 as commercial institutions with more than 700 million active users per month. He must ask for permission from Meta Before using the models. The open source attempt said that such a restriction çıkış out of the category of AI.open source”Although he claims the opposite of Meta.