News

Meta Unveils Llama 4 AI Models, Boasting Major Performance Gains and Open-Source Access

• By John K. Waters
• 04/07/2025

Meta Platforms has launched a new suite of AI models under its Llama 4 banner, expanding its capabilities in generative artificial intelligence and taking a direct swing at competitors such as OpenAI, Google, and xAI.

The company announced four new Llama models: Llama Scout, Maverick, Behemoth, and Reasoning, each tailored for distinct computational and application needs. The models introduce substantial increases in parameter size and new efficiency mechanisms, which Meta claims deliver best-in-class performance across a range of AI benchmarks.

"Each of these models serves a different purpose," said Meta's Content and Social Media Manager Andrew Hutchinson, in a blog post, "with Meta releasing variable options that can be run with less or more powerful systems. So, if you're looking to build your own AI system, you can use Llama Scout, the smallest of the models, which can run on a single GPU. "

Expanded Architecture, Tiered Access
Llama 4 Scout and Maverick, each equipped with 17 billion parameters, are designed to offer lightweight deployment options, including single-GPU operation. Both leverage a "mixture of experts" system that activates only subsets of model parameters per query — a move that reduces latency and compute requirements without sacrificing accuracy.

At the other end of the spectrum, the Llama 4 Behemoth model boasts more than 2 trillion parameters, making it, according to Meta, the largest publicly available AI model. While details on the Llama 4 Reasoning model remain sparse, Meta suggests it will target complex inference and decision-making tasks.

The parameter count represents more than a twofold increase over the company's previous Llama 3 generation, marking a steep progression in system reasoning depth. "Over time, Meta's building in more system logic to ask more questions, and dig further into the context," Hutchinson said.

Computational Firepower and Market Positioning
Meta's latest models are underpinned by vast hardware resources. The company now operates approximately 350,000 Nvidia H100 chips — a figure that significantly outpaces the estimated 200,000 used by OpenAI and Elon Musk's xAI. Meta is also accelerating development of its own AI chips to reduce dependency on third-party silicon.

That computational dominance positions Meta to lead in high-volume AI workloads, both for internal applications and external adoption.

"In terms of tangible, available compute and resources, Meta is pretty clearly in the lead," Hutchinson said.

Open Source, Strategic Distribution
A key element of the launch is Meta's decision to open source all Llama 4 models, allowing third-party developers to integrate them into their own AI systems. This strategy, already attracting major platforms like LinkedIn and Pinterest, is poised to increase the tech giant's influence in the broader AI ecosystem.

"External developers…stand to make Meta the key load-bearing foundation for many AI projects," Hutchinson said.

The models are expected to be integrated into Meta's own product suite in the coming weeks, enhancing chatbot experiences across WhatsApp, Facebook, Instagram, and Messenger. The improvements will also affect Meta's advertising infrastructure, powering smarter targeting and content generation.

Skepticism and Benchmark Transparency
While Meta has shared comparative performance charts, Hutchinson acknowledged that some observers have questioned the transparency and applicability of the company's internal benchmarks. The ultimate verdict on Llama 4's capabilities, he noted, will depend on real-world testing and user experience.

Despite these reservations, Meta claims that the Llama 4 models deliver superior results at a lower cost and barrier to entry than competing systems — a pitch likely to resonate in a developer community increasingly constrained by compute budgets.

Looking Ahead
Meta's Llama 4 rollout comes amid intensifying competition in the AI sector, with OpenAI's GPT-5 reportedly in development and Google expanding its Gemini model lineup. But with its open-source approach, massive compute capacity, and vertical integration across apps and ads, Meta is still in the race.