Photo credit: www.androidcentral.com
What you need to know
Meta is introducing two new models in its Llama 4 series—Scout and Maverick—both showcasing impressive capabilities in early evaluations. Scout excels in managing extensive documents, addressing intricate requests, and navigating large codebases, while Maverick adeptly integrates both textual and visual elements, making it particularly suited for smart assistants and conversational interfaces. These models are accessible now on Llama.com and through collaborations with platforms such as Hugging Face. Additionally, Meta AI is set to enhance user experiences across WhatsApp, Messenger, and Instagram in 40 countries, albeit initially limited to U.S. and English users.
With the launch of the Llama 4 series, Meta presents two distinct models: the streamlined Scout and the versatile Maverick. According to the company, early assessments indicate that these new models outperform rivals across various metrics.
Maverick is designed as a multifaceted tool, effectively processing both text and visual information, making it an excellent fit for applications like chat interfaces and digital assistants. In contrast, Scout is streamlined for tasks requiring enhanced focus, perfect for dealing with complex documents and extensive coding challenges.
The Llama 4 models are now available on Llama.com and through several of Meta’s partners, including Hugging Face. These new models are also being integrated into the Meta AI assistant, which is progressively launching across messaging platforms such as WhatsApp, Messenger, and Instagram in 40 nations. However, it’s worth noting that the multimodal capabilities are currently restricted to English-speaking users in the U.S.
MoE power
Meta has emphasized that Llama 4 represents significant advancements behind the scenes, marking the company’s initial utilization of a Mixture of Experts (MoE) configuration. This innovative approach enhances operational efficiency by subdividing substantial tasks into smaller, manageable segments, which are then delegated to specialized mini-networks.
Scout comprises 17 billion active parameters distributed across 16 distinct expert modules. Meta’s specifications suggest that this architecture outperforms Google’s Gemma 3, Gemini 2.0 Flash-Lite, and the open-source Mistral 3.1 in various standard benchmarks, all while maintaining a compact design that allows it to operate smoothly using just one Nvidia H100 GPU.
Scout particularly excels in synthesizing large volumes of text and making logical deductions within extensive codebases. A notable feature is its expansive context window, capable of handling up to 10 million tokens, making it adept at managing both textual and visual data on a grand scale.
Maverick’s strength
Similarly, Meta attributes strong performance metrics to Maverick, stating that it competes effectively with OpenAI’s GPT-4 and Google’s Gemini 2.0 Flash. Significantly, Maverick achieves comparable outcomes to DeepSeek-V3 in coding and logical tasks with less than half the active parameters.
Maverick utilizes 17 billion parameters distributed among 128 specialized expert networks. However, when measured against Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 Sonnet, or OpenAI’s GPT-4.5 for peak performance, Maverick does not maintain the same level of competitive edge.
Meta has also hinted at the development of Llama 4 Behemoth, currently in training, which they anticipate will rank among the most advanced large language models (LLMs) available. Behemoth is expected to feature 288 billion active parameters spanning 16 experts, with an overall parameter count nearing two trillion. Early assessments indicate that Behemoth surpasses GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in STEM tasks, particularly in mathematics, although it still falls short of Google’s Gemini 2.5 Pro in holistic performance metrics.
In summary, the incorporation of these cutting-edge models within Meta’s applications signifies a leap forward in the intelligence and responsiveness of their AI offerings, promising users more precise interactions, enhanced image generation, and advertisements that are more aligned with their interests.
Source
www.androidcentral.com