Meta Unveils Llama 4 AI Models, Surpassing GPT-4o and Grok 3 in Performance
Introduction to Meta’s Latest AI Models: Llama 4
After a four-month hiatus, Meta has launched a new line of Llama 4 AI models that include Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. Unlike earlier versions that utilized dense models, these new models adopt an innovative architecture known as MoE (Mixture of Experts). This change facilitates multimodal capabilities, allowing these models to process various types of data effectively from the ground up.
Overview of the Llama 4 Models
Llama 4 Scout
The smallest model, Llama 4 Scout, boasts an impressive total of 109 billion parameters, featuring 16 experts. However, it only activates 17 billion parameters during operations. One of its standout features is an extensive context length of up to 10 million tokens. According to Meta, Llama 4 Scout vastly outperforms earlier models like Gemma 3, Mistral 3.1, and Gemini 2.0 Flash Lite.
Llama 4 Maverick
The Llama 4 Maverick is considerably more advanced, packing a total of 400 billion parameters and 128 experts, although it still activates only 17 billion parameters at any given time. This model offers better performance than its predecessor, the Scout, due to its specialized expert configurations. The context length for Maverick is 1 million tokens. Meta asserts that Llama 4 Maverick performs better than competitors such as OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash.
Performance Metrics
One of the remarkable achievements of Llama 4 Maverick is its ELO score of 1,417 on the LMArena leaderboard, positioning it just below Gemini 2.5 Pro. This score surpasses other models like Grok 3, GPT-4o, and GPT-4.5. It also shows comparable performance to the latest DeepSeek V3 model in reasoning and coding tasks, despite utilizing only half the active parameters.
The Behemoth Model
Meta has invested significant effort into developing the Llama 4 Behemoth model, which is currently still in training. This expansive model aims to be one of the largest, with a total of 2 trillion parameters, though only 288 billion parameters are actively utilized across 16 experts. Meta proudly claims that Llama 4 Behemoth outperforms other major AI models like GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro, especially on STEM benchmarks. As it is not a reasoning-focused model, further refinements will allow for superior performance in future reasoning tasks based on the Llama 4 framework.
Availability and Features
The rollout of the Llama 4 models has commenced, with Meta making them available across various platforms such as WhatsApp, Messenger, Instagram, and the Meta AI website. This launch is occurring in 40 countries, although the multimodal features are currently restricted to users in the United States.
Key Takeaways
- Diverse Range of Models: The Llama 4 series includes Scout, Maverick, and Behemoth, each designed to cater to different needs and functionalities.
- Advanced Architecture: Utilizing MoE technology, these models demonstrate enhanced performance and versatility.
- Compelling Performance Metrics: The models have recorded impressive scores in various benchmarks, showing potential superiority over notable competitors.
- Availability Across Platforms: The models are being made accessible through multiple Meta platforms, expanding user reach.
Meta’s Llama 4 models represent a significant step forward in AI technology, showcasing innovative design and capabilities that set a high standard for future developments in artificial intelligence.
