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 publicly available models in Meta's fourth-generation Llama series and continues the company's strategy of distributing AI models outside the closed, API-only approach used by competitors such as OpenAI and Anthropic.

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 a larger, higher-capacity model aimed at more demanding tasks requiring greater reasoning and language capability.

Both models are built on a mixture-of-experts architecture, a design approach in which different subsets of the model's parameters activate depending on the input, allowing for greater parameter counts without proportional increases in computational cost per inference. Meta has not disclosed the precise total parameter counts for either model publicly.

Meta released the models under a custom license that permits commercial use subject to certain conditions, consistent with the licensing terms applied to previous generations of the Llama series.

Background

Meta began releasing Llama models publicly in 2023 with the original Llama series, followed by Llama 2 later that year and Llama 3 in 2024. The Llama series has become one of the most widely downloaded open-weight model families in the industry, with Meta reporting hundreds of millions of downloads across its generations.

Meta's open-model strategy positions it differently from peers such as OpenAI, Google DeepMind, and Anthropic, all of which primarily distribute frontier models through paid API access rather than direct model weights. The approach has drawn both support from the open-source developer community and criticism from some AI safety researchers who argue that releasing powerful model weights without restrictions increases misuse risk.

The Llama 4 release comes during a period of rapid model releases across the industry. In the weeks preceding this announcement, OpenAI released GPT-4.1 and the o3 and o4-mini reasoning models, Google DeepMind released Gemini 2.5 Pro, and xAI opened API access to its Grok 3 model.

Availability and Access

Both Llama 4 Scout and Llama 4 Maverick are available for download through Meta's official Llama website and through the Hugging Face model hub, where the Llama series has historically accumulated substantial download volume. Developers can apply directly through Meta's access portal, subject to the terms of the Llama 4 community license.

Meta also confirmed that the models are available via Meta AI, the company's consumer-facing AI assistant product, and through integrations with Meta's family of applications including WhatsApp, Instagram, and Facebook.

What It Means in Practice

Open-weight models allow developers to run inference on their own infrastructure, fine-tune models on proprietary datasets, and deploy applications without sending data to a third-party API. This is a distinction that has made previous Llama models popular with enterprises operating under strict data residency or privacy requirements, as well as with academic researchers who require direct access to model weights for study.

The mixture-of-experts architecture used in both Scout and Maverick is consistent with approaches recently adopted by other leading labs, including Google DeepMind in Gemini 1.5 and Mistral AI in its Mixtral models, suggesting broader industry convergence on the design pattern for scaling model capability efficiently.

Meta has indicated that additional models in the Llama 4 family, including larger variants, are in development and are expected to be released in subsequent months.