Meta Platforms Inc. released its Llama 4 series of generative artificial intelligence models in April 2025, introducing open-weight, natively multimodal capabilities to its developer ecosystem. The new series includes the Scout and Maverick models, with a larger Behemoth model currently in training, marking an expansion in Meta's strategy for widely accessible AI.
The Llama 4 series comprises multiple models, with Scout featuring 17 billion active and 109 billion total parameters, alongside a 10 million token context window. Maverick, also with 17 billion active parameters, has 400 billion total parameters and a 1 million token context window. Behemoth, planned for future release, is expected to include 288 billion active and 2 trillion total parameters. These models are designed to process extensive input data, with Scout's context window reportedly equivalent to approximately 80 average novels and Maverick's to eight.
Meta states that all Llama 4 models underwent training on "large amounts of unlabeled text, image, and video data" to achieve "broad visual understanding," alongside training on 200 languages. Scout and Maverick utilize a "mixture-of-experts" (MoE) architecture, reported to enhance training and inference efficiency. The Llama 4 models support text, image, and video input for diverse assistive tasks such as coding, document summarization across 12 languages, and complex data analysis. Maverick is positioned as a generalist model, while Scout targets extensive workflows and Behemoth is slated for advanced research and STEM applications.
The Llama 4 Scout and Maverick models are accessible via Llama.com and partner platforms, including Hugging Face. Meta has partnered with over 25 entities, including AWS, Google Cloud, Microsoft Azure, Nvidia, Databricks, Groq, Dell, and Snowflake, to host and support Llama deployments. While Meta's business model does not involve selling direct access to its openly available models, the company has established revenue-sharing agreements with model hosts. In May 2025, Meta launched the "Llama for Startups" program, offering support and potential funding to companies adopting its Llama models.
Meta provides a suite of tools to enhance the safety and security of Llama model use, including Llama Guard for content moderation, Prompt Guard against prompt injection attacks, CyberSecEval for cybersecurity risk assessment, Llama Firewall for secure AI systems, and Code Shield for filtering insecure code. Despite these tools, reports indicate that multimodal features are primarily limited to English, and the models, like other generative AIs, may generate false or misleading information. Concerns have also been raised regarding training data, which included pirated e-books and user posts from Facebook and Instagram; however, a federal judge sided with Meta in a copyright lawsuit, ruling its use of copyrighted works for training fell under "fair use." Benchmarking on LiveCodeBench revealed Llama 4 Maverick scored 40% on competitive coding problems, trailing OpenAI's GPT-5 (85%) and xAI's Grok 4 Fast (83%).
