Meta has introduced three new models from its artificial intelligence (AI) software Llama 4, with the aim of securing a position for itself in the generative AI (GenAI) competition. This move comes as Facebook and Instagram’s parent company seeks to play an active role in this emerging field.
Currently, two versions of the models have been made available, though their most potent variant—the Llama 4—is yet to be released, with Meta suggesting it will debut sometime in the future. In his announcement, CEO Mark Zuckerberg mentioned that this anticipated version, referred to as the Llama 4 Behemoth, would stand out as “the top-performing basic model globally.”
Although that model is still undergoing training, below are the details of the other two AI models currently available for download.
Llama 4 Scout
Among the three Llama 4 models introduced, Scout is the most compact yet intended to be the fastest.
Scout boasts 109 billion total parameters, serving as the fundamental components of the AI system that influence and dictate how the model handles data and produces outcomes. A greater number of these allows for tackling more intricate tasks.
It includes a Mixture of Experts (MoE) system, similar to a group of specialists where each expert focuses on particular tasks, for instance, mathematics. The Scout comprises 16 experts.
It also has a 10 million token context window, which means that it can digest about 8 million English words and then summarise them.
Meta says that Scout can handle multi-document summarisation, parsing extensive user activity for personalised tasks, and reasoning over vast codebases.
Llama 4 Maverick
The bigger Maverick Llama 4 edition features 128 specialists and has a combined parameter count of 400 billion.
Meta
mentioned in a blog post that this is the leading multimodal model, which means it can handle various types of data — including text, video, audio, and images — all at once to produce output.
Meta stated that their model outperformed competitors such as GPT-40 and Gemini 2.0 across various tests including coding, reasoning, multilingual capabilities, handling long contexts, and interpreting images. Additionally, it demonstrated performance levels similar to those of the significantly bigger DeepSeek v3.1 when tested for coding and reasoning skills.
Llama 4 Behemoth
Although it hasn’t been released yet, Meta stated that Behemoth boasts around 16 specialists and approximately two trillion overall parameters.
The firm stated that it provides cutting-edge performance for non-reasoning models in areas such as mathematics, multilingual capabilities, and image tests.
However, as reported by VentureBeat, when various reasoning models were employed to evaluate Behemoth alongside DeepSeek’s R1 and OpenAI’s O1 models, Meta’s solution did not consistently come out on top in every scenario.
Open weight
In his announcement video posted on Instagram, Zuckerberg stated that Meta aims to develop the world’s premier artificial intelligence system, release it as open-source software, and ensure global accessibility so that people everywhere can benefit from it. He has long believed that open-source AI will emerge as the dominant approach, and he sees initiatives like Llama 4 paving the way for this shift.
Nevertheless, the firm has referred to the Llama 4 models as open-weight rather than open-source following criticisms that they do not align with the official definition provided by the Open Source Initiative.
This primarily stems from Meta not disclosing the datasets used to train their AI models.
Is Llama version 4 accessible in Europe?
It is not available to customers based in Europe, which includes not being available for research or personal use.
Meta has not provided reasons for this, but
stopped the release of earlier Llama models
In the European Union, this is due to regulatory concerns since the training data for its models originates from Meta users.
However, if you reside outside of Europe, you have the option to download the models from either Meta or Hugging Face.
Changing political favour
Meta states that all Llama 4 models demonstrate significant enhancement regarding “political bias.”
The firm stated that “[particularly, leading large luggage models] have traditionally aligned with left-leaning views regarding contentious political and social issues.” They added that Llama 4 aims to be more accommodating towards right-wing perspectives, aligning with Zuckerberg’s support for former U.S. President Donald Trump following his victory in the 2024 elections.
Leave a Comment