Meta has gained a firm foothold in the generative AI market with its customizable open models. Now, with its new Llama 4 family, the social media giant is seeking to broaden its appeal to enterprises with powerful, natively multimodal models that are free or relatively affordable.
Meta's three new models are Llama 4 Maverick with 400 billion total parameters and Llama 4 Scout with 109 billion total parameters, both generally available now, and Llama 4 Behemoth, in preview. The two smaller models can be run on a single GPU, according to Meta.
The DeepSeek factor
In some ways, the release of the Meta Llama 4 line on April 5 is a response to the challenge posed to U.S. and European generative AI vendors by upstart Chinese generative AI vendor DeepSeek with its cheap, capable models.
All the new Meta models use a mixture-of-experts architecture -- in which subsets of a model are trained in specific subjects -- that is at the core of DeepSeek's models. The new Meta models are also priced to compete with the paid versions of DeepSeek, noted Andy Thurai, founder of advisory firm The Field CTO.
DeepSeek basically came up with a model that's cheaper, faster, more efficient and available for free. Meta actually wanted to beat that.
Andy ThuraiFounder, The Field CTO
"DeepSeek basically came up with a model that's cheaper, faster, more efficient and available for free," Thurai said. "Meta actually wanted to beat that."
The Llama 4 models, like their predecessors, are open weight as opposed to fully open source. With open weight models, vendors release the trained model parameters, or weights, but not the source code or training data.
Meta's Llama 4 models are also available in free and more powerful paid versions, and all are multimodal, capable of handling text, video and image inputs and producing text, video and image outputs. Some of DeepSeek's models are multimodal, while others are text-only.
As for Behemoth, with 2 trillion total parameters and 16 experts, it is intended to be used for distillation, a process in which a larger model teaches, or trains, smaller models.
"It is the biggest model ever built," Thurai said.
Meta and enterprises
With its previous Llama models, Meta found an audience among smaller and midsize enterprises that wanted to fine-tune models for marketing and e-commerce on Meta's social media platforms: Facebook, Instagram and WhatsApp. That suited Meta, which realized a larger economic benefit from a wider customer base, rather than seeking to generate revenue strictly from selling its models.
With its more powerful new models, Meta can appeal to larger enterprises with more advanced generative AI applications, said Arun Chandrasekaran, a Gartner analyst.
"Now, for example, a customer could be a manufacturing plant where you're trying to do predictive maintenance or object detection," he said. "Or you have products that are coming on the manufacturing shop floor, and you want to do things like product quality detection."
As for the DeepSeek comparison, Chandrasekaran said while the Chinese startup is a competitor, Meta "definitely has a lot more mindshare in this space."
"They have been able to create very capable open weight models for a long period of time," he added. "The combination of that, having multiple multimodal releases as well as the fact that they've signaled their intent to remain open weight in the foreseeable future puts Meta in a slightly different league when compared to, say, a DeepSeek or others like it."
Similarly, Meta's AI technology is an established presence in enterprises, said Mark Beccue, an analyst at Enterprise Strategy Group, now part of Omdia.
The open model market
However, Beccue also noted that Meta is facing increased competition in the open weight and open source generative AI market from the likes of not only DeepSeek, but also IBM and AWS. Others in the open source arena include the Allen Institute for AI and Mistral.
"Meta has done great with open source, and they do have an advantage in open source in that a lot of enterprises have been using Llama previously, and they're comfortable with it," Beccue said.
In regard to Meta's performance claims, Beccue said generative AI vendors are now in a phase of continually leapfrogging each other in benchmarking tests of model size, speed and intelligence. Any advantage is momentary.
"This is a supercharged Space Race. Things are morphing so quickly," he said.
According to Meta's comparisons, pricing per 1 million input and output tokens for Llama 4 Maverick, for example, is $0.19 to $0.49, while Google Gemini 2.0 Flash is $0.17, DeepSeek V3.1 is $0.48, and OpenAI's GPT-4o is $4.38.
Informa TechTarget AI news writer Esther Shittu contributed to this story.
Shaun Sutner is senior news director for Informa TechTarget's information management team, driving coverage of artificial intelligence, analytics and data management technologies, and big tech and federal regulation. He is a veteran journalist with more than 30 years of news experience.
