Mark Zuckerberg says Meta’s latest Llama models put open-source AI in the driving seat – News

Meta Platforms Inc. unveiled the latest additions to its Llama family of large language artificial intelligence models on Saturday, claiming that they’re among the most powerful ever released to the public.

The new models, which are part of the Llama 4 series, are available to access now through the Meta AI assistant on the web and in Messenger, WhatsApp and Instagram, and can also be downloaded from Meta itself or Hugging Face.

They include Llama 4 Scout, which is said to be a small model that fits inside a single Nvidia H100 graphics processing unit, and Llama 4 Maverick, which is a larger model more comparable OpenAI’s GPT-4o and Google LLC’s Gemini 2.0 Flash. Both of the models are said to have 17 billion active parameters.

The company is still working on the largest model within the Llama 4 lineup. According to Meta Chief Executive Mark Zuckerberg, it’s called Llama 4 Behemoth, and it will be the “highest performing base model in the world” once it’s released.

Meta says the Llama 4 models are the most advanced it has developed thus far, and also “best in class” in terms of their modality. Multimodal AI models are able to process different kinds of data formats, including text, images, audio and video, so they can comprehend more complex scenarios and generate better responses.

Meta Chief Product Officer Chris Cox told CNBC in March that the Llama 4 models are designed to power so-called AI agents, which are more sophisticated AI models and systems with enhanced reasoning skills and the ability to surf the web and take actions. They can be told to complete various tasks on behalf of humans, and they’ll do them with minimal supervision.

Meta said Llama 4 Scout is a small yet extremely efficient model that’s designed to run on just one high-end graphics card. It’s capable of processing up to 10 million “tokens,” which is the AI industry’s term for chunks of words or data. That represents a massive jump from previous “small” LLMs, the company noted.

Llama 4 Scout outperformed other small LLMs like Google’s Gemma 3 and Gemini 2.0 and Mistral’s Mistral 3.1 in a number of key benchmarks spanning a “broad range” of applications.

As for Llama 4 Maverick, that’s a more powerful iteration that’s designed to handle tasks such as writing code, creative writing, tackling math problems and understanding images and video. According to Meta, it outperforms rival models such as OpenAI’s GPT-4o and Google’s Gemini 2.0, even though it’s more efficient and cost-effective. It did not compare it with the recently released Gemini 2.5 model, but it said Maverick’s performance is also on a par with DeepSeek Ltd.’s V3 reasoning model, despite using less than half of its active parameters.

In a detailed blog post describing the new models and how they were created, Meta explained that it used a newer kind of system called “Mixture-of-Experts” or MoE, which allows them to work more efficiently. Rather than using the entire model for each task, MoE systems only activate the part needed to complete the task in hand, so they can run faster and use less energy.

Meta also talked a little about the upcoming Llama 4 Behemoth model, which will have 288 billion active parameters, and almost 2 trillion parameters in total when it launches. It’s still being trained, hence it hasn’t been released yet, but it’s already being used to “teach” the smaller Llama 4 models using a technique called “distillation,” which enables knowledge to be transferred from larger to smaller models. According to Meta, early tests show that Llama 4 Behemoth significantly outperformed competing models such as GPT-4.5 and Claude Sonnet 3.7 on a number of STEM benchmarks.

In addition to performance, Meta also focused on making the Llama 4 models safer and more balanced. It has enhanced the built-in protections that aim to prevent them from providing harmful or biased responses, so they can provide more balanced answers to controversial and politically sensitive questions. As such, the Llama 4 models will be less likely to refuse to answer tough questions, or lean too heavily on one side of the political spectrum, Meta said.

“Our goal is to build the world’s leading AI, open source it, and make it universally accessible so that everyone in the world benefits,” Zuckerberg said in a video on Instagram. “I’ve said for a while that I think that open source AI is going to become the leading models, and with Llama 4 this is starting to happen. Meta AI is getting a big upgrade today.”

We can expect to see much more from Meta on the AI front when it kicks off its first annual LlamaCon AI conference on April 29, when it may well release the Llama 4 Behemoth model. It’s also expected to announce a standalone Meta AI application at the event.

Image: News/Freepik AI Suite