Meta Launches Llama 4 Maverick With 128 Experts

Meta has officially released Llama 4 Maverick, the latest addition to its open-weight large language model family, featuring an unprecedented 128-expert mixture-of-experts (MoE) architecture that pushes the boundaries of efficient AI inference. The model represents a dramatic leap from previous Llama generations and positions Meta as a formidable challenger to proprietary models from OpenAI, Google, and Anthropic.

Maverick arrives alongside Llama 4 Scout, a smaller sibling in the Llama 4 lineup, but it is Maverick that has captured the AI community's attention with its sheer architectural ambition. The model activates only a fraction of its total parameters during inference, delivering performance that rivals — and in some benchmarks surpasses — models like GPT-4o and Claude 3.5 Sonnet, while remaining fully open-weight under Meta's community license.

Key Facts at a Glance

• Architecture: 128-expert MoE design with approximately 400 billion total parameters, but only ~17 billion active per forward pass
• Context window: Supports up to 1 million tokens in extended configurations
• Multimodal: Natively handles text and image inputs out of the box
• License: Released under Meta's updated community license, free for most commercial use
• Benchmarks: Outperforms GPT-4o on several reasoning and coding benchmarks according to Meta's internal evaluations
• Availability: Accessible via Hugging Face, Meta's own platform, and major cloud providers including AWS, Azure, and Google Cloud

Inside the 128-Expert MoE Architecture

The headline innovation in Llama 4 Maverick is its mixture-of-experts design, which uses 128 specialized sub-networks — called 'experts' — within the model. During each inference step, a gating mechanism selects only a small subset of these experts to process a given token, dramatically reducing the computational cost compared to a dense model of equivalent total size.

This approach is not entirely new — Google's Switch Transformer and Mixtral from Mistral AI both popularized MoE designs. However, Maverick takes the concept to a new scale. Where Mixtral 8x7B uses 8 experts, Maverick deploys 128, creating a vastly more granular routing system that can specialize in an enormous range of tasks and domains.

The practical benefit is significant. Despite having roughly 400 billion total parameters, Maverick activates approximately 17 billion during any single forward pass. This means the model can run on substantially less hardware than a comparably performing dense model, lowering the barrier to deployment for enterprises and independent developers alike.

Benchmark Performance Challenges Proprietary Leaders

Meta's internal benchmark results paint an impressive picture of Maverick's capabilities. According to the company's published evaluations, the model achieves competitive or superior scores across a range of standard benchmarks compared to leading proprietary models.

Key performance highlights include:

• MMLU (Massive Multitask Language Understanding): Maverick scores above 85%, placing it in the same tier as GPT-4o and Gemini 1.5 Pro
• HumanEval (coding): The model achieves a pass rate exceeding 80%, outperforming several closed-source competitors
• Reasoning tasks: On benchmarks like ARC-Challenge and HellaSwag, Maverick shows strong gains over Llama 3.1 405B
• Multimodal understanding: Early tests on image-text tasks show Maverick competing with dedicated vision-language models

Independent verification of these claims is still underway. The open-source community has already begun running Maverick through third-party evaluation suites, and early reports from researchers on social media broadly corroborate Meta's claims, though some note that performance can vary depending on quantization and hardware configuration.

It is worth noting that benchmark performance does not always translate directly to real-world utility. However, the sheer breadth of tasks where Maverick performs competitively suggests that Meta's architectural bet on a high-expert-count MoE design is paying dividends.

Multimodal Capabilities Expand Llama's Reach

Llama 4 Maverick is natively multimodal, a first for the Llama family at this scale. Previous Llama models required community-built extensions or fine-tunes to handle image inputs, but Maverick processes both text and images as first-class inputs from the start.

This multimodal capability opens up a wide range of applications. Developers can now build systems that analyze charts, interpret photographs, process documents with embedded images, and handle visual question-answering tasks — all using a single unified model.

Meta has trained Maverick on a diverse dataset that includes paired image-text data, enabling the model to reason about visual content with surprising nuance. Early demonstrations show the model accurately describing complex scenes, reading text within images, and performing spatial reasoning tasks that trip up many competing models.

The move to native multimodality also signals Meta's strategic direction. As the AI industry shifts toward unified models that handle multiple data types, Maverick ensures that the open-source ecosystem does not fall behind proprietary alternatives from OpenAI (GPT-4o) and Google (Gemini) that already offer robust multimodal support.

What This Means for Developers and Businesses

The release of Maverick carries major practical implications for the developer community and enterprise AI teams. The model's open-weight nature means organizations can download, fine-tune, and deploy it on their own infrastructure without per-token API costs.

For startups and mid-size companies, this is particularly impactful. Running a model with GPT-4o-class performance on private servers eliminates ongoing API expenses that can scale into tens of thousands of dollars per month for high-volume applications. The MoE architecture's efficiency further reduces hardware requirements, making deployment feasible on clusters with as few as 2 to 4 high-end GPUs using quantized versions.

Key use cases that developers are already exploring include:

• Enterprise chatbots and customer service agents with proprietary data kept on-premises
• Code generation and review tools integrated into CI/CD pipelines
• Document analysis systems leveraging multimodal capabilities for legal, financial, and medical documents
• Research assistants that combine text reasoning with image interpretation
• Edge deployment scenarios using heavily quantized versions of the model

Meta's community license permits commercial use for organizations with fewer than 700 million monthly active users, which effectively covers all but the largest tech companies. This licensing approach has been a cornerstone of Meta's open-model strategy, and it remains unchanged with the Llama 4 release.

Industry Context: Meta Doubles Down on Open AI

Maverick's release reinforces Meta's position as the leading proponent of open-weight AI models among major tech companies. While OpenAI, Google, and Anthropic continue to keep their most powerful models behind proprietary APIs, Meta has consistently argued that open models drive faster innovation and broader adoption.

Mark Zuckerberg has repeatedly framed open-source AI as a strategic priority for Meta. The company benefits from a thriving ecosystem of developers building on Llama models, which in turn strengthens Meta's AI infrastructure, attracts talent, and generates goodwill in the developer community.

The competitive landscape has intensified considerably. Mistral AI continues to iterate on its own MoE models, Google recently released Gemma 2 as an open alternative, and Chinese labs like DeepSeek have made waves with their own high-performing open models. Maverick's 128-expert architecture appears designed to leapfrog all of these competitors in one decisive move.

The timing is also notable. With regulatory scrutiny of AI increasing in both the US and EU, Meta's open approach positions the company favorably in policy discussions. Open models allow external auditing, academic research, and community-driven safety testing — arguments that Meta has leveraged effectively in conversations with lawmakers.

Looking Ahead: What Comes Next for Llama 4

Meta has hinted that Maverick and Scout are not the final entries in the Llama 4 family. Industry observers expect a larger model — potentially called Llama 4 Behemoth — to arrive later in 2025, possibly featuring even more experts and greater total parameter counts.

The broader trajectory is clear. Meta is betting that scale, efficiency, and openness can together create a flywheel that keeps Llama at the frontier of AI capabilities. If Maverick's early reception is any indication, that bet is paying off.

For developers, the immediate next step is clear: download, experiment, and benchmark. Maverick is available now on Hugging Face, Meta's Llama website, and through major cloud providers. With community fine-tunes and optimizations already appearing within hours of launch, the open-source ecosystem is moving fast.

The release of Llama 4 Maverick marks a pivotal moment in the open AI movement. A 128-expert MoE model performing at the level of the best proprietary systems — available for free download — was nearly unimaginable 2 years ago. Today, it is reality, and the implications for the AI industry will unfold for months to come.