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 for researchers and developers to download and deploy. The release marks the first public availability of Meta's fourth-generation Llama architecture, which the company says incorporates a mixture-of-experts design and supports natively multimodal inputs including text and images.

Model Specifications

Llama 4 Scout is a 17-billion active parameter model built on a mixture-of-experts architecture with 16 experts total. Meta states Scout supports a context window of up to 10 million tokens, which the company describes as the longest context window of any publicly available model in its class. Llama 4 Maverick is also a 17-billion active parameter model but draws on 128 experts, making it a larger and more computationally intensive system. Meta reports that Maverick matches or exceeds the performance of competing models including GPT-4o and Gemini 2.0 Flash on a range of standard benchmarks, while Scout is positioned as an efficiency-focused option suited to single-GPU deployment.

Both models were trained on a dataset Meta describes as comprising 22 trillion tokens, using up to 32 rows of Meta's custom-built Grand Teton GPU clusters during the training process.

Background

Meta has released successive generations of its Llama model family since 2023, each made available under a custom license that permits broad use including commercial applications, subject to usage restrictions for deployments above a defined user threshold. The Llama series has been widely adopted by enterprise developers, academic researchers, and AI startups as a foundation for fine-tuned and customised applications.

The Llama 4 generation follows Llama 3, which Meta released in 2024 and which included variants ranging from 8 billion to 405 billion parameters. The move to a mixture-of-experts architecture in Llama 4 represents a structural change from the dense transformer design used in prior Llama generations. Mixture-of-experts systems activate only a subset of parameters for each input, allowing models to maintain large total parameter counts while reducing per-inference compute requirements.

Meta has not released a Llama 4 model at the scale of Llama 3's 405-billion parameter variant with this initial drop. The company has indicated that additional Llama 4 models, including a larger system referred to as Llama 4 Behemoth, remain in development.

Availability and Licensing

Llama 4 Scout and Maverick are available for download through Meta's official Llama website and through the Hugging Face model repository. Meta has published model weights, model cards, and associated documentation alongside the release. The models are released under Meta's Llama 4 Community License, which follows the structure of prior Llama licenses with restrictions on use by services exceeding 700 million monthly active users, a threshold that effectively excludes the largest competing platforms from unrestricted use of the weights.

Meta is also serving both models through its own Meta AI assistant product and through its API, giving developers the option to access the models via inference endpoint rather than local deployment.

What It Means in Practice

Developers with access to a single high-end GPU can run Llama 4 Scout locally due to its active parameter footprint, according to Meta's technical documentation. Maverick requires more substantial infrastructure but remains within reach of multi-GPU setups common in enterprise and research environments. Both models accept image inputs natively without requiring separate vision modules, which expands their utility for applications involving documents, screenshots, and visual data.

Meta's continued open-weight release strategy keeps the company's models in active use across the developer community at a time when several competitors, including OpenAI and Anthropic, distribute their most capable models exclusively through paid API access.

Meta has indicated that Llama 4 Behemoth, a substantially larger model still in training, is expected to be released at a later date.