Meta Llama 4 Shakes Up the AI Landscape with Open-Source MoE Models

Efe Udin

April 7, 2025

Meta recently released the Llama 4 model, a new series built on the Mixture of Experts (MoE) architecture. According to experts, this model will redefine the limits of open-source AI. The launch marks a major shift in the AI market as it is marked to rival top systems like GPT-4o and Gemini 2.0. However, Llama 4 has a crucial edge: It is open source.

The lineup includes three key models, including the Llama 4 Scout, which is a light yet powerful multimodal model. There is also the Llama 4 Maverick, a high-performance rival that reportedly surpasses DeepSeek-V3 in benchmarks. Then, there is also the Llama 4 Behemoth, a staggering 2-trillion-parameter model that could outperform GPT-4.5. The company is yet to officially release the Llama 4 Behemoth.

Power Meets Efficiency

The computational efficiency of Llama 4 is among its most remarkable features. Llama 4 Llama 4 and Maverick both operate on a single H100 GPU, though with advanced features. This makes them cheaper than their close rivals. Meta provides API pricing that is lower than its rivals’, which could change how companies use AI on a large scale. According to Meta, though Maverick halves the number of parameters relative to its rivals, it does not compromise performance.

Benchmark Dominance and Open-Source Leadership

The Llama 4 Maverick has already caused a stir by beating the DeepSeek-V3 n benchmark. On the Grand Model Arena leaderboard, this model has a score of 1417 and is now the leading open-source model in coding, reasoning, and creative writing.

Meta says that this model will be accessible to developers globally. The Llama 4 series, in contrast to earlier Llama models, is naturally multimodal. This means that it can easily process text, images, and even tough visual reasoning tasks. For instance, the Scout advances open-source AI vision by being able to recognise objects, analyse images, and respond to contextual queries.

Read Also: Google to Develop Android OS Privately Starting Next Week

There is also an increase in the number of languages that this model supports, up to 200. This means that it will have a wider reach relative to previous models. In addition to raising the standard for open-source AI with Llama 4, Meta has increased pressure on rivals like Google and OpenAI.

Sam Altman, the CEO of OpenAI, has since given hints at the upcoming GPT. Meta’s new pricing and performance may force a major change in the AI behemoth’s handling of accessibility and pricing. We might soon witness a world in which open-source models drive the AI revolution.

Disclaimer: We may be compensated by some of the companies whose products we talk about, but our articles and reviews are always our honest opinions. For more details, you can check out our editorial guidelines and learn about how we use affiliate links.

Follow Gizchina.com on Google News for news and updates in the technology sector.

Source/VIA :

ITHome