April 08, 2025
Meta Unveils New AI Models as Llama 4 Behemoth Teased by Zuckerberg
Meta has introduced three new artificial intelligence models under its Llama 4 lineup, signaling a bold move in the generative AI race. While two of the models are already available for download, the most powerful version — dubbed the “Llama 4 Behemoth” — is still in training.
CEO Mark Zuckerberg called it “the highest performing base model in the world,” hinting at major capabilities yet to be seen. In the meantime, the released models offer a glimpse into Meta’s advancing AI technology and ambitions to compete with leading players in the GenAI space.
Llama 4 Scout:
Of the three Llama 4 models unveiled by Meta, Scout is the smallest—but also the fastest.
Scout features 109 billion parameters, which serve as the fundamental units that help the model understand and generate complex outputs. The more parameters a model has, the more sophisticated its capabilities.
It also uses a Mixture of Experts (MoE) architecture, with 16 experts, each specializing in specific tasks like mathematical reasoning or code analysis.
Scout boasts a 10 million token context window, enabling it to process and summarise the equivalent of around 8 million English words.
Meta says Scout is built for speed and efficiency, making it ideal for multi-document summarisation, analyzing large user activity datasets for personalization, and reasoning across complex codebases.
Llama 4 Maverick
The larger Maverick model from Llama 4 comes equipped with 128 experts and a massive 400 billion parameters.
According to Meta, Maverick is a best-in-class multimodal model, capable of processing and understanding multiple types of input—text, video, audio, and images—at the same time to generate seamless outputs.
In a blog post, Meta stated that Maverick outperforms comparable models like GPT-4o and Gemini 2.0 across key areas including coding, reasoning, multilingual understanding, long-context processing, and image interpretation.
It also rivals the much larger DeepSeek v3.1, particularly in coding and reasoning tasks.
Llama 4 Behemoth
Although it hasn’t been released yet, Meta revealed that Behemoth, the most powerful model in the Llama 4 lineup, will feature 16 experts and nearly 2 trillion total parameters.
Meta claims Behemoth delivers state-of-the-art performance in non-reasoning tasks, excelling in math, multilingual capabilities, and image benchmarks.
However, according to VentureBeat, when tested using various reasoning models, Behemoth did not consistently outperform competitors like DeepSeek’s R1 and OpenAI’s o1, suggesting that its dominance may depend on the specific task or benchmark.
Open weight
In his Instagram announcement video, Mark Zuckerberg stated that Meta’s mission is to “build the world’s leading AI, open source it, and make it universally accessible so that everyone in the world benefits.” He added, “I’ve said for a while that I think open source AI is going to become the leading models, and with Llama 4, that is starting to happen.”
However, despite those claims, Meta has labeled the Llama 4 models as “open weight” rather than fully open source. This distinction comes after criticism that the release doesn’t align with the Open Source Initiative’s official definition, mainly because Meta has not disclosed the training data used to develop the models.
Is Llama 4 available in Europe?
The Llama 4 models are currently unavailable to users in Europe, including for research or personal use. While Meta hasn’t officially explained why, it’s likely due to regulatory concerns, similar to the delays faced with previous Llama models. Those earlier setbacks were linked to the use of Meta user data in training the models, raising compliance issues under EU data protection laws.
However, for those outside of Europe, the models are available for download via Meta’s platform or through Hugging Face.
Changing political favour
Meta claims that all Llama 4 models demonstrate significant improvements in reducing political bias. According to the company, previous leading large language models have tended to lean left on controversial political and social issues, but Llama 4 aims to be more balanced—showing greater inclusivity toward right-leaning perspectives. This update aligns with Mark Zuckerberg’s recent outreach to U.S. President Donald Trump following his 2024 election victory.
(Source)