- Llama 4 adopts Mixture of Experts architecture, departing from Llama 3's dense transformer design.
- Scout model offers the largest context window of any openly available model at 10 million tokens.
- Maverick achieves 85.5% MMLU, 88.0% HumanEval, and 84.7% Math, competing with GPT-4o and Claude 3.5 Sonnet.
- Llama 4 Community License permits commercial use except for companies with over 700 million monthly active users.
- Maverick processes vision and text natively from training rather than as an add-on capability.
- Quantized Maverick runs on four A100 GPUs, eliminating per-token costs for organizations with scale.
What Is Llama 4?
Llama 4 is Meta AI's fourth generation of open-weightOpen-WeightA model whose trained weights are publicly available for download, allowing anyone to run, fine-tune, or build on top of it — distinct from fully open-source (which also includes training code and data).Learn more → large language models, released in April 2025. It represents a significant architectural leap from Llama 3. Where Llama 3 used a dense transformerTransformerThe neural network architecture that underpins virtually all modern LLMs, introduced in 2017, built around self-attention mechanisms that process entire sequences in parallel.Learn more →, Llama 4 adopts Mixture of ExpertsMixture of ExpertsA neural network architecture where only a subset of model parameters are activated for each input token, enabling very large models that are surprisingly efficient to run.Learn more → (MoE) architecture and adds native multimodalMultimodalCapable of processing and generating multiple types of data — such as text, images, audio, and video — within a single model.Learn more → capabilities.
The Llama 4 family consists of three models: Scout (the efficient option with a 10M context windowContext WindowThe maximum amount of text (measured in tokens) that a model can read and reason over in a single interaction, including your prompt, any documents, and previous conversation history.Learn more →), Maverick (the flagship multimodal model), and Behemoth (the largest, most capable model, released under the Llama 4 Community License for most uses).
Llama 4 Scout: The Long Context Marvel
Scout is the most technically remarkable member of the family. Its 10 million-tokenTokenThe basic unit of text that an LLM processes — roughly corresponding to a word or word-piece. Models read input and produce output in tokens, which is also how API usage is measured and billed.Learn more → context window is the largest of any openly available model. It enables processing entire large codebases, books, or years of conversation history in a single promptPromptThe input text sent to a language model — the question, instruction, or context that triggers a response.Learn more →. It uses a 16-expert MoE architecture that makes it surprisingly efficient for its capability level.
At $0.05/$0.17 per million tokens, Scout represents extraordinary value for long-document analysis, multi-file code understanding, and large-scale RAGRAGRetrieval-Augmented Generation — a technique that enhances LLM responses by first retrieving relevant documents from an external knowledge base and including them in the prompt.Learn more → applications. Its main limitation is that very long contexts can exhibit the 'lost in the middle' problem for retrieval-heavy tasks.
Llama 4 Maverick: The Open Frontier
Maverick uses 128 experts and processes up to 1M tokens. It competes credibly with GPTGPTGenerative Pre-trained Transformer — the model architecture and family name behind OpenAI's most famous models, from GPT-2 to GPT-5.Learn more →-4o and Claude 3.5 Sonnet across most benchmarks, a remarkable achievement for an open-weight model. MMLUMMLUA comprehensive benchmark evaluating language models across 57 academic subjects from elementary to professional level using multiple-choice questions.Learn more →: 85.5%, HumanEvalHumanEvalA benchmark dataset of 164 Python programming problems used to evaluate code generation capabilities of language models through unit test pass rates.Learn more →: 88.0%, Math: 84.7%. It is natively multimodal, enabling vision and text understanding from training rather than as a bolt-on.
Being open-weight means Maverick can be fine-tuned, quantized, and self-hosted. For organizations that need frontier-quality performance with data sovereignty, Maverick is the strongest option currently available. It has already spawned dozens of fine-tuned variants on Hugging Face within weeks of release.
The Llama 4 License
Llama 4 is released under the Llama 4 Community License, which permits commercial use for most organizations. The key restriction: companies with over 700 million monthly active users must contact Meta directly for a license. Competing AI service providers are also restricted from using Llama 4 as a backend without permission.
For the vast majority of developers and organizations, the license is effectively permissive. You can fine-tune, deploy, and build products on top of Llama 4 without royalties. This makes Llama 4 viable for serious commercial applications.
Llama 4 vs Closed Models: How Does It Stack Up?
Maverick is genuinely competitive with GPT-4o and Claude 3.5 Sonnet on most tasks, with some meaningful gaps. Llama 4 is weaker on complex multi-step reasoning, non-English languages, and precise instruction-following at the tail. For 70 to 80 percent of real-world tasks, the difference is negligible.
The open-weight advantage is real. Self-hosted Maverick with 4-bit quantizationQuantizationA compression technique that reduces model size and inference cost by storing weights in lower-precision numerical formats, trading a small amount of accuracy for much lower memory and compute requirements.Learn more → runs on four A100 GPUs and eliminates per-token costs. Organizations with significant scale and sensitivity requirements will find this equation compelling. The tradeoff is infrastructure complexity and the absence of OpenAI and Anthropic-level safety filtering.