Meta rolls out Llama 4 as race for AI dominance heats up

At launch, Llama 4 will offer two AI models, namely Llama 4 Scout and Llama 4 Maverick. One can download these models through Llama's website or Hugging Face. Llama 4 will also power Meta AI, the company's AI assistant, on the web and Meta-owned apps such as WhatsApp, Instagram, and Messenger.

Meta also previewed Llama 4 Behemoth, which it called "one of the smartest LLMs in the world and our most powerful yet to serve as a teacher for our new models".

The company stated that these are its first set of models built using a mixture-of-experts (MoE) architecture, which has attracted renewed attention due to the sudden rise of the Chinese AI lab DeepSeek.

Under this approach, developers split a large AI system into smaller sub-networks (or "experts"), each specialising in specific domains such as computer programming, physics, poetry, or biology. For any given input, only a subset of these experts is activated, enabling the model to process information in a more efficient manner.

This architecture enables large-scale models to significantly reduce computation costs during pre-training and enable faster performance during inference.

Read: AI pioneer Yann LeCun: India must embrace open source, invest in research to become an AI hub like France

Llama 4 Scout is a small AI model with 17 billion parameters and 16 experts, offering a context window of 10 million tokens. Zuckerberg said the model is designed to run on a single graphics processing unit (GPU). In March, Google also unveiled Gemma 3, a collection of lightweight open models that can run on a single GPU.

On the other hand, Llama 4 Maverick is a general-purpose model with 17 billion parameters and 128 experts. Meta says this model can run on a single host and is designed to be a "product workhorse model for general assistant and chat use cases."

The company claimed that this model outperforms OpenAI's GPT-4o and Google's Gemini 2.0 Flash on coding, reasoning, multilingual, long-context, and image benchmarks and is competitive with the much larger DeepSeek v3.1 on coding and reasoning benchmarks.

These launches come against the backdrop of DeepSeek claiming to have built AI models that can rival top-tier models from US companies such as Meta, Google, and OpenAI earlier this year. The move sparked fresh concerns among investors over the billions of dollars being poured into AI development by tech firms. However, top executives at Meta and Google have shrugged off any potential competition from DeepSeek in recent months.

Meta also stated that Llama 4 Behemoth, an AI model that is still in the training phase, will be a 288 billion active parameter model with 16 experts and nearly two trillion total parameters.

The social networking giant claimed that the model outperforms OpenAI's GPT4.5, Anthropic's Claude Sonnet 3.7, and Google's Gemini 2.0 Pro on several STEM (Science, Technology, Engineering, and Mathematics) benchmarks.

Reasoning AI model soon

Zuckerberg also announced a reasoning model called 'Llama 4 Reasoning' without disclosing much details. He noted that the company will share more details in the next month.

"This is just the beginning for the Llama 4 collection. We believe that the most intelligent systems need to be capable of taking generalised actions, conversing naturally with humans, and working through challenging problems they haven’t seen before," the company said in a blog post.

A fortnight ago, Meta had announced that Llama had crossed more than a billion downloads, up from 650 million downloads as of early December 2024.

In January, Zuckerberg announced that Meta will spend $60 billion to $65 billion in capital expenditure in 2025 to bolster its AI efforts. This includes investment into building infrastructure, including servers, data centers, and other infrastructure.