Meta unveils Llama 4: A new era of open AI models

Meta has unveiled the next generation of its open AI models — the Llama 4 series. According to internal tests, the models outperform competitors in several benchmarks, especially in STEM-related tasks.
 

The flagship of the series, Llama 4 Behemoth, is a large language model (LLM) with 2 trillion parameters that is still in the training phase. However, its multimodal «students» — Maverick and Scout — are already available to developers and users.
 

The updated Meta AI assistant, integrated into WhatsApp, Messenger, and Instagram, now runs on Llama 4 and is available in 40 countries. Multimodal features, however, are currently limited to the U.S.
 

New Generation Architecture

Llama 4 is the first Meta model series to use a Mixture of Experts (MoE) architecture. Maverick includes 128 experts and 400 billion parameters, with only 17 billion actively used. Scout features 16 experts, 109 billion parameters, and the same 17 billion active.
 

Internal testing showed that Maverick outperforms models like GPT-4o and Gemini 2.0 in areas such as coding, reasoning, long-context understanding, and image tasks. However, it still falls short of newer models like Gemini 2.5 Pro, Claude 3.7 Sonnet, and GPT-4.5.
 

Maverick is better suited for general assistant and chatbot applications, while Scout excels at document summarization and reasoning over large datasets. Scout can run on a single Nvidia H100 GPU, whereas Maverick requires a full Nvidia H100 DGX system or equivalent.
 

Controversy and Denial

After Maverick ranked second in LLM Arena — a test where users compare models and form a «user ranking» — several researchers pointed out that a special version of the model was used, one not available to the public. This version generated longer responses and used more emojis, raising concerns about real-world performance.
 

Meta’s VP of Generative AI, Ahmad Al-Dahle, denied that the model had been fine-tuned specifically for benchmarks. He explained that the varying quality users observed was due to the stabilization process of the final version.
 

«This is just the beginning of the Llama 4 series,» Meta stated. «We believe that intelligent systems must be capable of general reasoning, natural communication, and solving complex tasks they’ve never seen before. Empowering Llama with such capabilities will help us build better products across our platforms and unlock new innovation for developers.»