Meta Launches Llama 4 Scout and Maverick Models

Meta has officially released Llama 4 Scout and Llama 4 Maverick, the first two models in its next-generation Llama 4 family, introducing a groundbreaking 10 million token context window that dwarfs anything currently available from competitors. The release marks a significant leap in open-weight AI development and positions Meta as a formidable challenger to closed-source leaders like OpenAI and Google DeepMind.

The models represent Meta's most ambitious AI release to date, combining a novel mixture-of-experts (MoE) architecture with unprecedented context length capabilities. Both models are available immediately through Meta's ecosystem and popular AI platforms.

Key Takeaways at a Glance

• Llama 4 Scout features 17 billion active parameters with 16 experts, totaling 109 billion parameters
• Llama 4 Maverick scales up to 17 billion active parameters with 128 experts, totaling 400 billion parameters
• Both models support a 10 million token context window — roughly 7.5 million words or 15 full-length novels
• Scout fits on a single NVIDIA H100 GPU node when quantized to INT4
• Maverick delivers performance competitive with GPT-4o and Gemini 2.0 Flash on major benchmarks
• Both models are released under Meta's open-weight license, enabling commercial use

Mixture-of-Experts Architecture Powers Efficiency Gains

Mixture-of-experts is the architectural innovation driving Llama 4's performance-per-compute advantage. Rather than activating every parameter for each token processed, MoE models selectively route inputs to specialized 'expert' sub-networks. This means Llama 4 Scout only activates 17 billion of its 109 billion total parameters at any given time.

This approach delivers a dramatic efficiency improvement over dense transformer models. Scout can run on hardware configurations that would be insufficient for a traditional 109 billion parameter model, making it accessible to a far wider range of developers and organizations.

Maverick takes the same concept further with 128 experts. Despite its 400 billion total parameter count, it maintains the same 17 billion active parameter footprint during inference. This design choice allows Maverick to punch well above its computational weight class, competing with models that require significantly more resources to deploy.

Meta trained both models using a massive multilingual and multimodal dataset, with early support for image understanding alongside text. The company has indicated that native multimodal capabilities — including video and audio — will expand in future Llama 4 releases.

The 10 Million Token Context Window Changes the Game

The 10 million token context window is perhaps the most headline-grabbing feature of the Llama 4 family. To put this in perspective, here is how it compares to current industry offerings:

• GPT-4o (OpenAI): 128,000 tokens
• Claude 3.5 Sonnet (Anthropic): 200,000 tokens
• Gemini 1.5 Pro (Google): 2 million tokens
• Llama 3.1 (Meta, previous gen): 128,000 tokens
• Llama 4 Scout/Maverick: 10 million tokens

At 10 million tokens, Llama 4 can ingest and reason over entire codebases, full legal document repositories, or years of corporate communications in a single prompt. This capability has been a long-standing goal in the AI industry, as longer context windows reduce the need for complex retrieval-augmented generation (RAG) pipelines and chunking strategies.

Meta reportedly achieved this breakthrough through innovations in positional encoding and attention mechanisms, combined with training on extremely long-sequence data. The company claims the models maintain coherent reasoning across the full context length, though independent benchmarks from the research community are still forthcoming.

Benchmark Performance Rivals Closed-Source Giants

Meta has published benchmark results showing that both models perform competitively with leading proprietary alternatives. Llama 4 Maverick matches or exceeds GPT-4o on several widely-tracked evaluations, including MMLU, HumanEval, and reasoning-focused benchmarks.

Scout, despite being the smaller model, reportedly outperforms Llama 3.1 70B and Gemma 2 27B across the board. Meta positions it as the ideal model for developers who need strong performance within tighter compute budgets.

Key benchmark highlights include:

• MMLU (Massive Multitask Language Understanding): Maverick scores competitively with GPT-4o, significantly above Llama 3.1 405B
• HumanEval (Code Generation): Both models show improved code generation over Llama 3 series
• Reasoning tasks: Substantial improvements on mathematical and logical reasoning benchmarks
• Multilingual performance: Enhanced capabilities across 12+ languages including Spanish, French, German, Hindi, and Japanese
• Multimodal understanding: Early image comprehension capabilities rival specialized vision-language models

It is worth noting that Meta's self-reported benchmarks should be taken with appropriate caution until third-party evaluations confirm the results. The AI community has grown increasingly skeptical of vendor-published benchmarks, particularly after several high-profile discrepancies in 2024.

What This Means for Developers and Businesses

The practical implications of Llama 4's release are substantial for multiple stakeholder groups. For developers, the open-weight nature means full access to model weights for fine-tuning, distillation, and custom deployment. Unlike OpenAI's GPT-4o or Anthropic's Claude, there are no API-only restrictions.

For enterprises, the 10 million token context window could eliminate entire categories of infrastructure complexity. Companies that have invested heavily in RAG systems to work around context limitations may find that Llama 4's native long-context capabilities offer a simpler alternative. Document analysis, legal review, and codebase understanding use cases stand to benefit immediately.

For AI startups, the availability of a competitive open-weight model at this performance tier lowers the barrier to entry significantly. Teams that previously relied on expensive API calls to OpenAI or Anthropic can now self-host comparable capabilities, potentially reducing inference costs by 50-80% depending on deployment scale.

The hardware accessibility is equally noteworthy. Scout's ability to run on a single H100 node with INT4 quantization means that cloud deployment costs could be remarkably low compared to running a 400 billion parameter dense model.

Meta's Open-Weight Strategy Intensifies AI Competition

Meta's decision to continue its open-weight strategy with Llama 4 sends a clear signal to the industry. CEO Mark Zuckerberg has repeatedly argued that open-source AI development creates a healthier ecosystem and ultimately benefits Meta's core social media and metaverse businesses.

This release puts additional competitive pressure on several fronts:

• OpenAI faces growing competition from models that are free to use and modify
• Google sees its Gemini context window advantage significantly challenged
• Anthropic must contend with an open alternative that approaches Claude-level performance
• Mistral and other open-weight competitors face a formidable new rival from a company with virtually unlimited compute resources

The broader trend is clear: the gap between open-weight and closed-source models continues to narrow. Each Llama generation has closed the distance faster than the previous one, and Llama 4 appears to represent the closest parity yet with frontier closed-source systems.

Looking Ahead: Llama 4 Behemoth and Beyond

Meta has confirmed that a third model in the Llama 4 family — codenamed Llama 4 Behemoth — is still in training. Behemoth is expected to be significantly larger than Maverick and is being positioned as Meta's answer to the most capable frontier models from OpenAI and Google.

The company has also hinted at expanding Llama 4's multimodal capabilities in coming months, with native video understanding and audio processing on the roadmap. These additions would transform Llama 4 from a primarily text-focused model family into a truly universal AI system.

Several key questions remain for the AI community to evaluate in the coming weeks:

• How well does the 10 million token context window perform on real-world long-document tasks versus synthetic benchmarks?
• What are the actual inference costs when deploying Maverick at scale?
• How quickly will the fine-tuning community produce specialized variants for coding, medical, legal, and other vertical applications?
• Will Behemoth maintain the open-weight license, or will Meta restrict access to its most powerful model?

For now, Llama 4 Scout and Maverick represent a major milestone in the democratization of frontier AI capabilities. Developers can access both models through Hugging Face, Meta's own platforms, and major cloud providers including AWS, Google Cloud, and Microsoft Azure. The models are available immediately, with Meta providing comprehensive documentation and deployment guides.

The AI landscape continues to evolve at a breakneck pace, and Meta's latest release ensures that the open-weight movement remains a central force shaping the industry's future. Whether you are building a startup, deploying enterprise AI, or conducting research, Llama 4 demands attention as one of the most significant model releases of 2025.