In a significant step to expand its influence in the rapidly evolving world of generative AI, Meta has officially launched the first set of models under its latest open-source suite, Llama 4. The announcement was made on Saturday, accompanied by a video from CEO Mark Zuckerberg on Instagram, where he reaffirmed the company’s commitment to making AI universally accessible through open-source development.
“Our goal is to build the world’s leading AI, open source it, and make it universally accessible,” Zuckerberg stated. “I’ve said for a while that open-source AI will lead the way, and with Llama 4, we’re starting to see that happen.”
The initial release includes two models — Llama 4 Scout and Llama 4 Maverick — both of which are now available for download through Meta’s official Llama website and the AI platform Hugging Face. These models also power Meta AI, the company’s advanced virtual assistant integrated into services like WhatsApp, Instagram, Messenger, and the web.
Meta also introduced Llama 4 Behemoth, calling it the most powerful model it has developed to date. Still in development, Behemoth is expected to play a critical role in training and shaping future models. Its architecture is built on a mixture-of-experts (MoE) framework, marking Meta’s first implementation of this approach. MoE divides the model into specialized expert modules, each trained in a specific domain such as physics, poetry, biology, or programming. During any task, only the most relevant experts are activated, making the system more efficient and cost-effective in both training and inference.
Llama 4 Scout is designed to be lightweight and highly efficient, featuring 17 billion parameters and 16 expert modules. With a 10-million-token context window and the ability to run on a single GPU, it caters to developers seeking powerful models without needing extensive hardware. Llama 4 Maverick, also equipped with 17 billion parameters but with 128 experts, is crafted as a general-purpose model aimed at assistant-style applications. Meta describes Maverick as a reliable workhorse capable of handling a wide range of tasks, including conversation, reasoning, code generation, and multilingual understanding.
According to Meta, Maverick outperforms several top-tier models, including OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash, across multiple benchmarks. These include areas such as long-context reasoning, code generation, and image-based analysis. It is also said to rival DeepSeek v3.1, a significantly larger model, in tasks requiring logical reasoning and programming skills.
Llama 4 Behemoth is still under development but is expected to feature 288 billion active parameters, 16 experts, and nearly two trillion total parameters. Meta claims the model already surpasses GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in STEM-related evaluations. The company is also working on another upcoming model in the lineup — Llama 4 Reasoning — which will be tailored specifically for advanced problem-solving and analytical tasks. More information on this model is expected to be released in the coming weeks.
The Llama 4 launch comes amid growing global competition in AI. Earlier this year, Chinese company DeepSeek made headlines by claiming that its models matched or exceeded the performance of leading U.S. offerings. While Meta and Google have downplayed these claims, the development signals a broader race in the AI space that is only intensifying.
Meta also shared that downloads of Llama models recently surpassed one billion — a sharp increase from 650 million in December 2024. This surge in adoption underscores the growing popularity of open-source AI tools and the demand for accessible, high-performing models.
To support its expanding AI ambitions, Meta previously announced that its infrastructure spending for 2025 will range between $60 billion and $65 billion. This investment will go toward servers, data centers, and other essential infrastructure to scale its AI initiatives.
With the release of Llama 4, Meta has not only advanced the state of open-source AI but also set the stage for a future where powerful language models are more accessible, efficient, and adaptable than ever before.