Meta Releases Llama 4 Scout and Maverick Open Models

What Happened

Meta released two new large language models, Llama 4 Scout and Llama 4 Maverick, making both available as open-weight models that developers and researchers can download and deploy. The release marks the first models in Meta's fourth-generation Llama series and continues the company's strategy of distributing capable AI models outside of a closed, subscription-based access model.

The Models

Llama 4 Scout is designed as a lightweight, efficient model intended for deployment in resource-constrained environments, including on-device and edge computing scenarios. Llama 4 Maverick is the larger of the two, targeting more demanding tasks that require greater reasoning and language capability.

Both models use a mixture-of-experts architecture, a design approach in which only a subset of the model's total parameters are activated for any given input. This architecture allows the models to maintain a large overall parameter count while keeping computational costs lower during inference than a comparable dense model would require.

Llama 4 Scout features 17 billion active parameters with 109 billion total parameters across 16 experts. Llama 4 Maverick also carries 17 billion active parameters but draws from a pool of 400 billion total parameters across 128 experts. Scout supports a context window of 10 million tokens. Maverick supports a context window of one million tokens.

Availability and Licensing

Meta released the model weights through its official Llama website and through Hugging Face, a widely used platform for distributing open-weight AI models. The models are available under Meta's custom Llama 4 Community License, which permits commercial use subject to conditions including a restriction on using the models to train competing large language models above a specified user threshold.

Meta also made both models accessible via its Meta AI assistant and through its AI Studio platform for developers building applications.

Background

Meta has released successive generations of Llama models since 2023, positioning the series as a direct alternative to closed models offered by OpenAI, Anthropic, and Google. The Llama 3 series, released in 2024, included models ranging from 8 billion to 405 billion parameters and was widely adopted by developers, enterprises, and academic researchers.

Meta does not charge for access to Llama model weights, though users are responsible for the infrastructure costs of running the models. The open-weight approach has contributed to broad third-party adoption, with Llama models integrated into cloud platforms including Amazon Web Services, Microsoft Azure, and Google Cloud.

The company has framed open-weight AI development as beneficial to the broader research community and as a counterweight to what it describes as excessive concentration in closed AI systems. Meta chief executive Mark Zuckerberg has made public statements advocating for open-source AI development on multiple occasions.

What It Means in Practice

Developers can download the Llama 4 Scout and Maverick weights and run them on their own infrastructure without routing queries through Meta's servers or paying per-token API fees. This allows organizations with data privacy requirements or cost constraints to deploy capable models internally.

The 10-million-token context window in Scout is among the largest publicly available in an open-weight model, enabling processing of very long documents or large codebases within a single inference call. Maverick's benchmark results, as reported by Meta, place it competitively against leading closed models on several standard evaluations including multilingual tasks and coding benchmarks.

Meta has not released a Llama 4 model at the scale of the Llama 3.1 405 billion parameter model, with the company indicating that larger Llama 4 models remain in development.

What Comes Next

Meta has indicated that additional models in the Llama 4 family, including larger variants, are planned for release later in 2025.