Llama 4 Maverick: Meta's open flagship

5 April 2025 · 10 min · Model Profiles

TL;DR Llama 4 Maverick is Meta's open multimodal MoE: 400B total, 17B active over 128 experts, with a native 1M context. It sits right in the middle of the strongest open weight tier, with the deepest tooling and fine-tuning ecosystem around it. The license is the custom Llama 4 Community License; big-user thresholds and attribution apply, so review before a consumer-scale product. A multi-GPU node serves it; a workstation cannot. Strong pick for RAG-heavy and document-intense internal systems.

Meta's Llama line has the most installed base of any open-weight family, and Llama 4 Maverick is the model keeping that crown. For teams that already run Llama infrastructure, Maverick is the natural upgrade path; for everyone else it is a proven, well-supported multimodal open model with a 1M context. This post goes through what it is, where it fits, and the two things to check before shipping: hardware and license.

What is Llama 4 Maverick?

Llama 4 Maverick is Meta's large open-weight release from the Llama 4 generation [model card](https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E) (P1): 400 billion total parameters, 17 billion active per token across 128 experts, early-fusion native multimodality for text and images, and a native context of 1 million tokens. Released April 5, 2025 [Meta announcement](https://ai.meta.com/blog) (P3), it has been integrated into more serving stacks, fine-tuning frameworks and deployed products than any other model in this survey.

What is it good at?

Maverick's two arguments are ecosystem and modality. First, ecosystem: the amount of battle-tested tooling, community tooling and fine-tuning know-how around Llama 4 dwarfs every other line, which lowers the risk and time of production adoption. Second, native multimodality with a 1M context makes it a strong pick for document-heavy RAG, image-understanding pipelines, and internal systems that mix text and screenshots [Artificial Analysis](https://artificialanalysis.ai) (P2), [BenchLM](https://benchlm.ai) (P2). It is a strong generalist, if not the single top scorer on every 2026 leaderboard.

Can you legally use it commercially?

The Llama 4 Community License is a custom agreement (P1). It permits commercial use, redistribution, attribution and naming requirements, and applies large-user thresholds in the hundreds-of-millions-of-MAU zone that can require additional permission from Meta, with some regional considerations noted in the documentation. Internal use and ordinary products are normally fine after a quick review; giant consumer-scale distribution needs time with a real licensing lawyer. The post on [what open-source AI means after the license shifts](https://dintechnologies.com/blog/what-open-source-ai-means-after-2025-license-shifts) explains where this sits versus MIT and Apache.

What hardware does it take?

Maverick is node-class:

• A multi-GPU node of high-memory GPUs is the realistic serving target.
• The active 17B keeps the per-token cost reasonable relative to the 400B total, but the total still dictates the memory required.
• Workstations and single GPUs are out of reach.

Sizing active-versus-total memory follows the standard MoE rules in [hardware tiers for local inference](https://dintechnologies.com/blog/2026-hardware-tiers-local-ai-inference).

What does it take to run it?

Because of the ecosystem, the path is unusually well travelled:

1. Pull a quantized build that fits your node. The trade-offs are in [quantization and memory optimization](https://dintechnologies.com/blog/2026-quantization-moe-efficiency-memory-optimization).

2. Serve with vLLM, which supports the Llama family deeply; the engine comparison is in [local AI runtimes compared](https://dintechnologies.com/blog/2026-inference-runtimes-local-ai-tooling-maturity).

3. Set a context budget per request; a 1M window is an explicit memory purchase.

4. Keep corpora and prompts inside your network. The argument is the same as [running open models privately](https://dintechnologies.com/blog/running-open-models-privately).

What does it cost?

The ownership rule from the [cost reality check](https://dintechnologies.com/blog/mid-year-cost-reality-check) applies cleanly. At sustained volume, especially multimodal or long-document volume, a private Maverick node replaces easily-metered API spend, and the ecosystem maturity reduces the hidden engineering cost of getting there.

When is Maverick the wrong choice?

• Your procurement policy demands MIT or Apache: take DeepSeek V4, GLM-5.2 or Gemma 4 instead.
• You need the top open reasoning scores of 2026: Kimi K3 and DeepSeek V4 Pro sit ahead.
• You need a 10M context for a giant corpus: that is literally the sibling post, on [Llama 4 Scout](https://dintechnologies.com/blog/llama-4-scout-profile).

Where this leaves you

Pick Maverick if you value a proven build, native multimodal, and the safest ecosystem adoption path in the industry. Run a small pilot with your exact multimodal and long-context prompts, confirm the license thresholds stay clearly under your product's reality, and then take advantage of the most supported open-weight deployment stack in existence. If the license tables turn you off, the MIT and Apache 2.0 equivalents are a documented alternative in this series.