Meta Releases Llama 4 Scout and Maverick Open Models

What Happened

Meta released two open-weight large language models, Llama 4 Scout and Llama 4 Maverick, making them available for download to developers and researchers. The release marks the first public availability of Meta's fourth-generation Llama model family, which the company says uses a mixture-of-experts architecture and supports natively multimodal inputs including text and images.

Model Specifications

Llama 4 Scout is a 17-billion active parameter model with 16 experts, designed for deployment on a single GPU. Meta states it supports a context window of up to 10 million tokens, which the company describes as among the longest available in any model at its size class. Llama 4 Maverick is also a 17-billion active parameter model but uses 128 experts, making it a larger and more capable system intended for higher-resource deployments. Meta says Maverick achieves competitive results on standard benchmarks against models including GPT-4o and Gemini 2.0 Flash. Both models were trained on a dataset Meta describes as comprising 22 trillion tokens, incorporating text and image data.

Architecture and Design

The mixture-of-experts design activates only a subset of parameters for any given input, a technique intended to improve computational efficiency relative to dense models of equivalent total parameter counts. Meta says the native multimodal capability was built into the models from the initial training stage rather than added through post-training adapters, a distinction the company highlights as differentiating Llama 4 from its predecessors. The 10-million token context window in Scout is designed to allow processing of very long documents or large codebases within a single inference call.

Availability and Licensing

Both models are available for download through Meta's official Llama website and through Hugging Face under Meta's custom community license. The license permits commercial use subject to certain conditions, consistent with the terms applied to prior Llama generations. Meta is also making the models available through major cloud providers including AWS, Google Cloud, Azure, and others, as well as through Groq and Cerebras for inference. The models are accessible via Meta AI, the company's consumer-facing assistant product, in supported regions.

Background

Meta first released the original Llama model in February 2023, initially restricting access to researchers before broadening availability. Subsequent releases, including Llama 2 in July 2023 and Llama 3 in April 2024, progressively expanded the scale and capability of the open-weight series. The Llama model family has become widely used in enterprise and research settings as one of the most prominent openly available large language model lineups. Meta has positioned open-weight model releases as central to its AI strategy, contrasting with the API-only access models offered by OpenAI and Anthropic. Llama 4 arrives as competition in the open-weight segment has intensified, with DeepSeek releasing high-performing open models in early 2025 that drew significant attention from developers globally.

What It Means in Practice

Developers can download Scout and Maverick weights and run them on local infrastructure or deploy them through cloud platforms without ongoing API fees, subject to the license terms. The extended context window in Scout allows tasks such as full-codebase analysis or processing of book-length documents in a single pass. Maverick, with its larger expert count, is positioned for use cases requiring higher accuracy where greater compute resources are available. Enterprise users accessing the models through cloud providers can integrate them into existing infrastructure using standard API interfaces.

What Comes Next

Meta has indicated that additional Llama 4 models, including a larger model referred to as Llama 4 Behemoth, are in development and expected to be released at a later date.