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 both available for download through Meta's official channels and partner platforms. The release marks the first generation of the Llama model family to natively support multimodal inputs, including text and images, and introduces a mixture-of-experts architecture to the Llama lineup.

Model Specifications

Llama 4 Scout is built on a mixture-of-experts architecture with 17 billion active parameters and 109 billion total parameters. Meta states it supports a context window of 10 million tokens, which the company describes as the longest context window of any openly available model at launch. Llama 4 Maverick also uses a mixture-of-experts design, with 17 billion active parameters and 400 billion total parameters.

Both models are natively multimodal, trained from the ground up to process interleaved text and image inputs rather than relying on a separately bolted-on vision component. Meta has made the model weights available under its Llama 4 Community License, which permits commercial use subject to terms outlined in the agreement.

Background

Meta has released successive generations of Llama models since 2023, with Llama 2 and Llama 3 each drawing significant adoption among researchers, enterprises, and developers who prefer open-weight models they can run and modify independently. The Llama 3 family, released in 2024, included models ranging from 8 billion to 405 billion parameters and established Meta's position as a primary supplier of open-weight models competitive with proprietary offerings.

The mixture-of-experts architecture used in Llama 4 is the same general approach employed by several other frontier models, including Mistral's Mixtral series and, according to public reporting, earlier versions of OpenAI's GPT-4. The design activates only a subset of total model parameters for any given input, reducing compute requirements relative to a dense model of equivalent total parameter count.

Meta Chief AI Scientist Yann LeCun and CEO Mark Zuckerberg have both stated publicly in prior interviews that Meta intends to keep Llama models openly available, framing open-weight releases as a strategic priority for the company.

Availability and Access

The models are available via Meta's website, Hugging Face, and a set of cloud and API partners including Amazon Web Services, Google Cloud, Microsoft Azure, and others. Meta also made the models accessible through its own Meta AI assistant products.

Llama 4 Scout, given its smaller active parameter footprint, is positioned by Meta for deployment on single-GPU hardware, while Maverick requires more substantial infrastructure. Meta has also disclosed that a third model in the Llama 4 family, Llama 4 Behemoth, remains in training and has not yet been released. Behemoth is described by the company as a teacher model used in the training of Scout and Maverick, with approximately 2 trillion total parameters.

Benchmark Performance

Meta published benchmark results alongside the release showing Scout and Maverick performing competitively against models including Google's Gemini 2.0 Flash and Anthropic's Claude 3.5 Sonnet on a range of standard evaluations covering reasoning, coding, and multimodal understanding tasks. Independent third-party verification of these results was not immediately available at the time of release.

What It Means in Practice

Developers and organizations using prior Llama models will need to review the updated Llama 4 Community License terms before deploying the new models in production environments, as licensing conditions have historically varied between Llama generations.

Meta has indicated that Llama 4 Behemoth will be released publicly at a later date once training is complete.