Meta's Llama 4 Brings Frontier AI to Self-Hosted Deployments

30-Second Summary

Meta's Llama 4 family has made frontier-class AI capability accessible to organisations that need full control over their data and infrastructure. With Scout fitting on a single NVIDIA H100 and Maverick available through AWS, Azure, and Snowflake within existing cloud agreements, the barriers to self-hosted AI have dropped significantly. For operators in finance, healthcare, or any sector with data sovereignty requirements, Llama 4 represents the clearest path yet to deploying a capable AI model without sending sensitive data to an external provider. Licensing is not fully open source and requires legal review, and some benchmark claims have been challenged by independent testing, but the cost and deployment story is compelling.

At a Glance

• Topic: Model Releases
• Company: Meta
• Date: 5 April 2025 (launch); broadly adopted across enterprise platforms through early 2026
• Announcement: Meta released Llama 4 Scout and Maverick as open-weight models under the Llama 4 Community Licence, with a third model (Behemoth) in limited preview
• What Changed: For the first time, frontier-class AI performance is available for self-hosted enterprise deployment at consumer-grade infrastructure cost
• Why It Matters: Organisations with data sovereignty requirements, regulated industry obligations, or cost constraints can now access near-frontier AI without a third-party cloud dependency
• Who Should Care: COOs, CIOs, compliance leads, and operators in finance, healthcare, legal, and defence

Key Facts

• Company: Meta
• Launch Date: 5 April 2025 (Scout and Maverick); Behemoth in limited preview, no broad release date confirmed
• What Changed: Open-weight frontier AI models available for self-hosted deployment, with dedicated regulated-industry documentation and major cloud platform integrations
• Who It Affects: Any organisation evaluating AI deployment with data sovereignty, cost, or compliance constraints
• Primary Source: Meta AI Blog, llama.com, official deployment documentation

What Happened

Meta released Llama 4 Scout and Maverick on 5 April 2025, introducing a new architecture class to the open-weight model landscape. Both models use a Mixture of Experts (MoE) design, where only a fraction of total parameters activate per inference, delivering high capability at low compute cost.

Llama 4 Scout carries 17 billion active parameters across 16 experts and supports a 10-million-token context window, the largest of any publicly available model at launch. This means Scout can process entire large codebases, lengthy legal contracts, or extensive conversation histories in a single pass. It fits on a single NVIDIA H100 GPU, making on-premises deployment practical for organisations that already run GPU infrastructure.

Llama 4 Maverick uses the same 17 billion active parameters but scales to 128 experts, for a total of 400 billion parameters. Its context window is 1 million tokens. This is the model Meta uses internally across Facebook, Instagram, and WhatsApp. It is available via AWS SageMaker JumpStart, Microsoft Azure AI Studio, Snowflake Cortex AI, GroqCloud, and Together AI, meaning organisations already operating in these environments can access Maverick within their existing security perimeters and without new vendor agreements.

Meta has published dedicated deployment guides for regulated industries at llama.com, covering finance, healthcare, and defence use cases with Kubernetes and vLLM configurations. Red Hat partnered with Meta for day-one production-grade vLLM support, signalling enterprise-readiness intent from the infrastructure layer.

A third model, Llama 4 Behemoth, was announced alongside Scout and Maverick with approximately 288 billion active parameters and 2 trillion total parameters. Behemoth remains in limited preview and is not broadly available.

Why It Matters

• Data sovereignty is no longer a blocker for frontier AI. Organisations in regulated industries can now deploy a capable model entirely within their own infrastructure, with no data leaving their environment
• The cost differential is material. Maverick runs at approximately 91 percent less per token than GPT-4o at comparable serving configurations, which changes the ROI calculation for any high-volume AI workflow
• Scout's 10-million-token context window enables document-heavy workflows that were impractical with smaller context models, including full contract review, codebase analysis, and extended research tasks
• Cloud integrations with AWS, Azure, and Snowflake mean organisations can access Llama 4 within existing procurement and security frameworks, without a new vendor evaluation cycle
• The MoE architecture delivers competitive benchmark performance while activating only a fraction of total parameters, keeping inference costs low even at scale
• Independent testing has identified gaps between advertised and real-world long-context performance, meaning thorough evaluation on your own data is required before committing to production deployment

The David and Goliath View

The most significant thing about Llama 4 is not its benchmark position. It is what it makes possible for organisations that have been sitting on the sideline because they cannot justify sending their most sensitive data to an external AI provider.

Until recently, the choice was binary: accept the data residency risk of a top-tier closed model, or accept the capability compromise of a smaller open-weight alternative. Llama 4 Scout and Maverick change that calculus. They are not the best models on every benchmark, but they are capable enough for the majority of enterprise workflows, they cost a fraction of closed alternatives, and they can run in your own environment with documented, production-grade deployment paths.

The licensing caveats are real. This is not OSI open source, and EU-based organisations face specific access restrictions. Any team treating Llama 4 as freely available software without legal review is taking on unnecessary risk. But for organisations that do the homework, the opportunity to run a frontier-class model in-house without sending data to Meta, OpenAI, or Anthropic is now a practical reality, not a theoretical one.

The recommendation is straightforward: if your organisation has avoided AI adoption because of data sovereignty or compliance concerns, Llama 4 removes your most defensible reason for waiting.

Where This Fits in the AI Stack

Secure AI Brain: Llama 4's self-hosting capability is the most direct expression of the Secure AI Brain model. Organisations can run a frontier-class model on their own infrastructure, with full control over data residency, access permissions, and audit trails. This is the deployment model that regulated industries have been waiting for.

Employee Amplification Systems: Scout's 10-million-token context window enables employees to work with entire document archives, codebases, or data histories in a single session, without the manual chunking and retrieval workarounds required by smaller-context models. This is a practical productivity gain for any knowledge-intensive role.

AI Growth Engine: The cost differential makes high-volume AI workflows economically viable for smaller organisations that previously could not justify the per-token cost of closed frontier models. Customer-facing automation, content generation, and data enrichment at scale become accessible at a different price point.

Questions Operators Are Asking

Is Llama 4 actually open source? No. It is released under the Llama 4 Community Licence, which permits downloading, self-hosting, fine-tuning, and commercial use in most cases. However, it is not OSI-certified open source. Companies with more than 700 million monthly active users require a separate licence from Meta. Organisations based in or operating primarily from the EU face access restrictions, likely due to AI Act and GDPR considerations. Have your legal team review the licence before production deployment.

How does the performance compare to GPT-4o or Claude? On many standard benchmarks, Maverick is competitive with GPT-4o and outperforms it on multimodal tasks. However, independent testing found real-world long-context performance significantly below Meta's advertised scores. Claude 3.5 Sonnet holds a meaningful advantage on coding evaluations. The honest answer is: it depends on your specific use case. Run your own evaluation on representative tasks before committing.

What infrastructure do we need to self-host Scout? Scout fits on a single NVIDIA H100 GPU. For production deployments, Meta recommends vLLM or TensorRT-LLM as inference engines. llama.cpp supports CPU-only and Apple Silicon deployments for lower-throughput use cases. Kubernetes deployment is documented and supported. Meta's regulated-industry documentation at llama.com covers finance, healthcare, and defence configurations specifically.

Can we access Llama 4 through our existing cloud contracts? Yes, in most cases. Maverick is available through AWS SageMaker JumpStart, Azure AI Studio, Snowflake Cortex AI, and GroqCloud. If your organisation already uses one of these platforms, you can likely access Maverick within your existing security perimeter and procurement framework without a new vendor agreement.

What about Llama 4 Behemoth? Behemoth was announced alongside Scout and Maverick but is not broadly available. It remains in limited preview. Meta's internal benchmarks show Behemoth outperforming GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on reasoning tasks, but these results have not been independently verified. Treat Behemoth as a directional signal for where open-weight models are heading, not as a production option today.

Citable Summary

What happened: Meta released Llama 4 Scout and Maverick, open-weight AI models using a Mixture of Experts architecture, available for self-hosting on standard enterprise GPU infrastructure and through AWS, Azure, and Snowflake within existing cloud agreements.

Why it matters: For the first time, organisations with data sovereignty requirements have a credible pathway to frontier-class AI that runs entirely within their own infrastructure, at approximately 91 percent lower per-token cost than GPT-4o.

David and Goliath view: Llama 4 removes the most defensible reason for regulated-industry operators to delay AI adoption. The licensing requires legal review and the benchmarks require independent validation, but the self-hosting pathway is real, documented, and commercially viable.

Offer relevance:

• Secure AI Brain: self-hosted frontier AI with full data residency control and no third-party dependency
• Employee Amplification Systems: 10-million-token context enables document-heavy workflows without chunking workarounds
• AI Growth Engine: 91 percent lower per-token cost makes high-volume AI workflows economically viable for smaller organisations