Meta Launches Llama 4 Scout and Maverick
Meta has officially rolled out the first models in its latest Llama 4 series, promising to transform personalized multimodal experiences. The two models, Llama 4 Scout and Llama 4 Maverick, are designed to deliver top-tier performance in processing various data types like text, images, video, and audio, all while maintaining efficiency and cost-effectiveness.
Llama 4 Scout, equipped with 17 billion active parameters and 16 experts, is being hailed as the leading multimodal model. It runs efficiently on a single NVIDIA H100 GPU and surpasses notable rivals like Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 across industry benchmarks. Additionally, it features an industry-leading 10 million token context window, enabling advanced tasks such as multi-document summarization and in-depth code analysis.
Llama 4 Maverick, meanwhile, also boasts 17 billion active parameters but utilizes 128 experts, providing exceptional performance across benchmarks, even outpacing GPT-4o and Gemini 2.0 Flash in reasoning and coding tests despite using fewer parameters than competitors. It offers an outstanding price-to-performance ratio, scoring an impressive ELO of 1417 on the LMArena experimental chat leaderboard.
Llama 4 Behemoth
Meta’s Llama 4 Scout and Maverick models leverage insights from the powerful, yet unreleased Llama 4 Behemoth, which already surpasses top models in STEM tasks. Committed to open innovation, Meta has made Scout and Maverick available for download on llama.com and Hugging Face, and integrated them into platforms like WhatsApp, Messenger, Instagram Direct, and Meta.AI.
A New Era for the Llama Ecosystem
Meta’s Llama 4 Scout and Maverick models mark a new chapter in its AI journey, using mixture-of-experts (MoE) architecture for greater efficiency. Maverick activates only parts of its 400B parameters, reducing latency and costs. Both models feature early fusion for seamless text-visual integration, expanded multilingual training across 200 languages, and the new MetaP method for optimized hyperparameters and better task performance.
Also read: Ant Group Taps Chinese Chips to Train AI Models, Cutting Costs and Reducing US Tech Dependence
Post-Training Excellence and Model Versatility
Beyond pre-training, Llama 4 Scout and Maverick undergo advanced post-training, including supervised fine-tuning, reinforcement learning, and direct preference optimization (DPO). Meta improved their reasoning, coding, and conversation abilities by focusing on challenging prompts and dynamically filtering data.
Llama 4 Maverick excels in image understanding and natural language tasks, outperforming rivals while being cost-effective. Meanwhile, Llama 4 Scout offers scalability with an impressive 10 million-token context length for complex tasks like code analysis and document summarization.
Both models use the iRoPE architecture, combining interleaved attention layers and adaptive scaling to enable exceptional length generalization and future support for nearly infinite context windows.
The tech giant teased further updates to the Llama ecosystem, promising deeper insights and future developments at LlamaCon on April 29.
