Meta Releases Llama 4 Models With Multimodal Capabilities

Meta released its Llama 4 family of large language models on April 5, 2025, marking the company's first open-weight models to natively process both text and images. The release introduces two immediately available models, Llama 4 Scout and Llama 4 Maverick, alongside a preview of a larger model called Llama 4 Behemoth, which remains in training.

What Happened

Meta published the Llama 4 Scout and Llama 4 Maverick models for download on its website and through platforms including Hugging Face. Both models use a mixture-of-experts architecture, a design in which only a subset of the model's total parameters are activated for any given input, allowing for lower computational costs at inference time compared to dense models of equivalent parameter counts.

Llama 4 Scout is described by Meta as a 17-billion active-parameter model with 16 experts, giving it a total parameter count of approximately 109 billion. Meta stated it fits within a single H100 GPU in certain configurations. Llama 4 Maverick is a 17-billion active-parameter model with 128 experts and a total parameter count of approximately 400 billion. Meta said Maverick achieves competitive benchmark scores against GPT-4o and Google's Gemini 2.0 Flash on several standard evaluation tasks.

Both models support a context window of up to 1 million tokens, which Meta said is the longest context window in the Llama series to date. The models accept image and text inputs and return text outputs.

Background

Meta began releasing models under the Llama name in February 2023, with subsequent generations following in 2023 and 2024. The Llama 3 series, released in 2024, became widely used in both research and commercial applications and was downloaded hundreds of millions of times, according to figures Meta has cited in public statements.

The Llama series is released under a licence that permits commercial use with certain restrictions, including a requirement that applications with more than 700 million monthly active users obtain a separate licence from Meta. This positions the models differently from fully closed systems offered by OpenAI, Anthropic, and Google, while also differing from permissive open-source licences that place no restrictions on use.

Mixture-of-experts architecture has been used in several recent high-profile model releases, including Mistral's Mixtral series and Google's Gemini 1.5 models, making the design an increasingly common approach among frontier model developers.

The Behemoth Preview

Meta also disclosed details about Llama 4 Behemoth, a model still under development. The company described it as having approximately 288 billion active parameters and nearly 2 trillion total parameters across its expert layers, which would make it one of the largest models disclosed publicly by any developer in terms of total parameter count. Meta provided benchmark figures suggesting Behemoth outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several reasoning and science benchmarks, though the model is not yet available for external use and the evaluation conditions were not independently verified at the time of publication.

Meta said Behemoth is being used as a teacher model to improve the quality of the smaller released models through a process called distillation.

What It Means in Practice

The release makes natively multimodal open-weight models more broadly accessible to developers and researchers who build on the Llama ecosystem. Prior Llama models processed text only, and developers seeking image understanding capabilities from open-weight models relied on separate vision components or third-party integrations.

The 1-million-token context window allows inputs that can include very long documents, extended conversation histories, or large volumes of images in a single request, a capability that has previously been associated primarily with closed commercial models.

Meta has stated that additional Llama 4 series models are planned for release later in 2025, with Behemoth expected to become available to external users following the completion of its training run.