Meta has released two groundbreaking open-weight AI models, Llama 4 Scout and Llama 4 Maverick, that fundamentally change what developers can build without relying on expensive proprietary APIs. Scout features an industry-leading 10-million-token context window, while Maverick competes directly with OpenAI's GPT-4o and Google's Gemini 2.0 Flash. Both models are natively multimodal, meaning they understand text, images, and video from the ground up, not as an afterthought. Most importantly, both are available as free, open-weight downloads on Hugging Face and llama.com .

What Makes These Models Different From Previous Llama Releases?

Llama 4 Scout and Maverick represent the first time Meta has built Llama models using a mixture-of-experts (MoE) architecture, a design pattern that activates only a portion of the model's parameters at any given time. This approach delivers powerful performance while keeping computational costs low. Scout contains 17 billion active parameters spread across 16 experts, with 109 billion total parameters. Maverick scales this up dramatically, maintaining the same 17 billion active parameters but distributing computation across 128 experts, with 400 billion total parameters .

The architectural shift matters because it allows these models to punch well above their weight. Scout can fit on a single NVIDIA H100 GPU with Int4 quantization, making it accessible to research labs, startups, and individual developers who lack expensive multi-node GPU clusters. Maverick runs on a single H100 DGX host or can be distributed across multiple nodes for higher throughput. Both models were pre-trained on over 30 trillion tokens, double the training data used for Llama 3, spanning 200 languages with over 100 languages represented by at least 1 billion tokens each .

How Can Developers Actually Use These Models?

• Long-context applications: Scout's 10-million-token context window enables full-codebase analysis, book-length document processing, and multi-session conversational memory, capabilities previously exclusive to API-only services.
• Fine-tuning and customization: Open-weight availability allows developers to download the full models, inspect them, fine-tune them for specific applications, and deploy them without API dependencies or vendor lock-in.
• Multimodal understanding: Both models process text, images, and video inputs using an early fusion architecture, where visual and textual information are processed jointly from the earliest layers rather than bolted on as separate modules.
• Multilingual deployment: With 200 languages supported, developers can build applications that serve global audiences without retraining or fine-tuning for language coverage.

How Do Scout and Maverick Compare to Competitors?

Scout outperforms Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 across a broad range of benchmarks, including coding, reasoning, long-context tasks, and image understanding . The 10-million-token context window has no equal among open-weight models, making it a unique advantage for applications requiring extensive document processing or conversation history.

Maverick positions itself as a direct competitor to GPT-4o and Gemini 2.0 Flash. Meta claims it beats both across widely reported benchmarks while achieving comparable results to DeepSeek V3 on reasoning and coding tasks at less than half the active parameters. On LMArena, a community-run benchmark leaderboard, Maverick achieves an ELO rating of 1417, placing it among the highest-performing open-weight models available .

The main limitation is that Llama 4 uses a community license rather than Apache 2.0, which imposes some restrictions on commercial use and redistribution. Google's competing Gemma 4, released just days earlier on April 2, uses the more permissive Apache 2.0 license. For organizations that prioritize licensing flexibility, this remains an important consideration .

What's the Bigger Picture for Open-Source AI?

Meta is not just releasing these models as downloads. Llama 4 is being deployed across Meta's consumer products, powering Meta AI on WhatsApp, Messenger, Instagram Direct, and the meta.ai website. This integration gives Llama 4 immediate access to Meta's billions of monthly active users, creating a deployment scale that no other open-weight model can match. For developers, this dual strategy means Meta benefits from community improvements and bug reports while using its own products as the largest testing ground .

The release follows months of anticipation after Meta announced it would invest $135 billion in AI infrastructure in 2026. Llama 4 represents the technical outcome of that investment: a model family designed to compete with proprietary offerings from Google, OpenAI, and Anthropic while remaining freely available to the developer community. Meta also disclosed Llama 4 Behemoth, a massive model with 288 billion active parameters across 16 experts and approximately 2 trillion total parameters, which is still training. Behemoth already outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM benchmarks and serves as the teacher model for knowledge distillation into Scout and Maverick .

For developers and organizations building on open models, Llama 4 Scout and Maverick set a new standard for what is available outside of API-only services. The combination of MoE architecture, native multimodality, record-breaking context windows, and massive multilingual coverage creates a model family that competes directly with the best proprietary offerings while remaining freely downloadable and customizable.