Meta has announced the release of “Llama 4,” featuring two new models aimed at competing with artificial intelligence giants such as OpenAI and Google. This article reveals the details of Meta’s artificial intelligence model and its anticipated capabilities.
Article contents:
- Introduction
- The new Llama 4 models
- MoE technology and its importance
- Competition and open-source licensing
- Conclusion
- Frequently Asked Questions
Introduction
On April 6, 2025, Wes Davis announced the launch of Meta’s latest version of Llama, under the name “Llama 4,” which includes two main models: “Llama 4 Scout” and “Llama 4 Maverick.” This comes as part of Meta’s efforts to establish itself as a fierce competitor against companies like OpenAI and its GPT-4o models, as well as Google and its Gemini 2.0 Flash model.
The new Llama 4 models
According to statements by Mark Zuckerberg, CEO of Meta, the company is currently working on another model named “Llama 4 Behemoth,” which some consider the largest project in terms of parameters and intensive training, describing it as “the highest-performing foundation model in the world.”
The “Llama 4 Scout” is relatively small in size, fitting on a single Nvidia H100 GPU graphics card. Despite its small size, this model can handle a context window of up to 10 million tokens, outperforming lightweight models like Google’s Gemma 3 and Gemini 2.0 Flash-Lite, as well as open-source models like Mistral 3.1.
As for the “Llama 4 Maverick,” it comes with greater capabilities that compete with models such as GPT-4o, Gemini 2.0 Flash, and DeepSeek-V3, consuming half of the active parameters needed by other models to reach the same level of performance. This move is an attempt by Meta to showcase the efficiency of the Mixture of Experts (MoE) approach it has adopted, which allocates a portion of the model to handle specific tasks, making the process more efficient.
Meta also pointed out that “Llama 4 Behemoth” comes with 288 billion active parameters, with total parameters reaching 2 trillion. Although this model has not yet been officially released, company data indicates its superiority in Science, Technology, Engineering, and Math (STEM) benchmarks over both GPT-4.5 and Claude Sonnet 3.7.
MoE technology and its importance
According to information from Meta, these models heavily utilize the Mixture of Experts (MoE) architecture, an innovative technique that divides the model into specialized units that handle the exact tasks needed, rather than running all parameters at once. Meta confirmed it will announce more details about the future of Meta’s artificial intelligence model during the LlamaCon conference scheduled for April 29.
Competition and open-source licensing
However, the “open-source” aspect of Llama 4 remains a subject of debate, because Meta imposes certain restrictions on commercial licenses for organizations with more than 700 million monthly active users, prompting critics like the Open Source Initiative to object to calling it fully “open source.”
Conclusion
If Meta’s promises of superior performance and higher efficiency in the Llama 4 models hold true, this could be a significant step toward competing with OpenAI and Google models. These models also open the door to wider applications in various fields, ranging from text content preparation and creation to programming, game development, and scientific research.
Frequently Asked Questions
1. What is new in Llama 4?
It includes two main models named Scout and Maverick, with a third massive model in preparation named Behemoth, all of which rely on MoE technologies.
2. What are the most prominent features of Llama 4 Scout?
Small size and the ability to run on a single Nvidia H100 GPU processing unit, with a context window of up to 10 million tokens.
3. How does Llama 4 Maverick compare to other models?
Meta indicates its superiority over GPT-4o and Gemini 2.0 Flash in performance and speed, while consuming fewer resources.
4. What is the status of Llama 4 Behemoth?
It is still in the development stage, featuring 288 billion active parameters and a total parameter count of 2 trillion, with promising performance in benchmarks.
5. Is the release completely open source?
Not entirely, as Meta’s license imposes some restrictions on commercial use by large companies, drawing criticism from the open-source community.
