In a strategic push to strengthen its foothold in the generative AI space, Meta released the first models from its latest open-source AI suite, Llama 4, on Saturday.
In an Instagram video, Meta CEO Mark Zuckerberg shared the company's bold AI ambitions: "Our goal is to build the world's leading AI, open source it, and make it universally accessible... 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 two initial models- Llama 4 Scout and Llama 4 Maverick-are now available for download via the Llama website and Hugging Face. These models also serve as the foundation for Meta AI, the company's virtual assistant now integrated across WhatsApp, Instagram, Messenger, and the web.
Additionally, Meta introduced Llama 4 Behemoth, describing it as one of the smartest large language models (LLMs) yet, and the most powerful version they've developed. It's intended to help train and guide future models.
This rollout also marks Meta's first use of a mixture-of-experts (MoE) framework. MoE divides the model into specialised components-each focused on areas like physics, poetry, biology, or programming. During any given task, only the most relevant expert modules are activated, boosting efficiency and reducing costs for training and inference.
Model Highlights
Llama 4 Scout is built with 17 billion parameters and 16 experts, offering a 10-million-token context window. Designed to operate on a single GPU, it reflects a trend toward lightweight, high-performance models-similar to Google's recent Gemma 3 release.
Llama 4 Maverick, also at 17 billion parameters but with 128 experts, is intended as a general-purpose model for a wide range of assistant-style use cases. Meta describes it as a reliable "workhorse" capable of handling chat, reasoning, and other digital tasks.
Meta claims Maverick outperforms OpenAI's GPT-4o and Google's Gemini 2.0 Flash across several benchmarks, including code generation, reasoning, multilingual understanding, image analysis, and handling long contexts. It also rivals DeepSeek v3.1, a much larger model, in performance across coding and logic-based tasks.
Earlier this year, DeepSeek claimed its models could match those from leading U.S. firms, prompting concerns over increasing global competition in AI. However, Meta and Google executives have downplayed its impact.
What's Next: Behemoth and Beyond
Still in development, the Llama 4 Behemoth model is expected to feature 288 billion active parameters, 16 experts, and nearly 2 trillion total parameters. Meta says it outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in STEM-related tests.
Zuckerberg also teased Llama 4 Reasoning, a model focused specifically on complex problem-solving and analytical tasks. More details are expected in the coming weeks.
"This is just the start of the Llama 4 lineup," Meta wrote in a blog post. "We believe the most advanced AI systems must be capable of taking generalized actions, engaging in natural conversations, and tackling problems they've never encountered before."
Scaling Up
Meta revealed that downloads of Llama models surpassed one billion just two weeks before Llama 4's launch- up from 650 million in December 2024.
Back in January, Zuckerberg announced the company's AI infrastructure spending for 2025 would range between $60 billion and $65 billion, covering investments in servers, data centres, and other resources needed to support Meta's expanding AI efforts.
