Llama 4 Maverick: 400B Total, 17B Active, and a Leaderboard Lesson

Llama 4 Maverick launched on 5 April 2025 with a striking headline: a leaderboard position just behind the strongest closed model of the time. Within a week that number had been qualified, and the model people could actually download was ranked far lower. Both facts are true, and understanding why is more useful than either on its own.

Maverick is a capable, efficient multimodal model. It is also a case study in reading model launches carefully. This guide covers both.

The architecture

Maverick is a mixture-of-experts model with about 400B total parameters and 17B active per token. Meta describes alternating dense and MoE layers; in the MoE layers each token goes to a shared expert plus one of 128 routed experts. Only a small slice of the network runs for any given token, which is why per-token compute is closer to a 17B dense model than to a 400B one.

It is natively multimodal through what Meta calls early fusion: text and image tokens enter the same backbone and were pretrained together, rather than bolting a vision encoder onto a finished language model. Input is text and images; output is text. Meta lists twelve supported languages (Arabic, English, French, German, Hindi, Indonesian, Italian, Portuguese, Spanish, Tagalog, Thai and Vietnamese), a 1M-token context window, roughly 22 trillion multimodal training tokens, and an August 2024 data cutoff.

Maverick was codistilled from Llama 4 Behemoth, a much larger teacher model that Meta previewed alongside it but never released. If the total-versus-active distinction is new to you, active vs total parameters explains why both numbers matter and for different reasons.

What the benchmarks show

Meta's published instruct results, beside its smaller sibling Scout:

• MMLU Pro — Maverick 80.5, Scout 74.3
• GPQA Diamond — 69.8, 57.2
• LiveCodeBench (Oct 2024 to Feb 2025) — 43.4, 32.8
• MMMU — 73.4, 69.4
• MathVista — 73.7, 70.7
• ChartQA — 90.0, 88.8
• DocVQA — 94.4, 94.4

The gap over Scout is widest on reasoning and code, where 128 experts and more total capacity pay off. On document and chart reading the two are essentially tied, which is worth remembering if images of documents are your main workload. At launch Meta positioned Maverick against GPT-4o and Gemini 2.0 Flash and as competitive with DeepSeek V3 at under half the active parameters. Those were 2025 comparisons; the frontier has moved a long way since.

The LMArena episode

Meta's launch post quoted an LMArena Elo of 1417 for "an experimental chat version" of Maverick. That entry, listed as Llama-4-Maverick-03-26-Experimental, had been tuned for conversational preference and was not the model Meta released. LMArena later said Meta should have made that clearer, published more than two thousand of its battles, and changed its submission policies. Those battles showed long, emoji-heavy answers. When the released model was scored, contemporary reports placed it 32nd, below older models such as Claude 3.5 Sonnet.

The lesson is not that the model is bad. It is that a preference leaderboard measures a specific, tunable property, and that a number from a variant you cannot run tells you nothing about the variant you can. Always confirm which checkpoint a score belongs to. Building your own small eval is the dependable alternative.

Running it

Meta released Maverick in BF16 and FP8. The FP8 weights fit on a single H100 DGX host, meaning an eight-GPU node, not a single card. That is the realistic self-hosting floor. All 400B parameters have to be resident even though only 17B are active, so memory, not compute, decides where it fits.

Through an API it is inexpensive for its capability. OpenRouter lists it at about $0.19 per million input tokens and $0.65 per million output at the time of writing. Context in practice depends on the host: at launch, providers served anything from roughly 500K up to the full 1M, so check the limit on the endpoint you use rather than assuming the model card figure.

The licence and the EU clause

Maverick ships under the Llama 4 Community License Agreement, a custom commercial licence. The familiar terms apply: companies above 700 million monthly active users need a separate licence from Meta, distributed products must display "Built with Llama", and models trained on Llama outputs must carry "Llama" at the start of their name.

The term that catches teams out is in the acceptable use policy. Meta's FAQ states that the rights to the Llama 4 multimodal models are not granted to individuals domiciled in the EU or companies with their principal place of business there, though end users of products built on them are not affected. Maverick is multimodal, so an EU-based company planning to self-host or fine-tune it needs legal review first. The licence comparison sets this beside MIT and Apache models that carry no such restriction.

Maverick or Scout, and what comes after

Pick Maverick when reasoning and code quality matter and you are calling an API or have a full GPU node. Pick Llama 4 Scout when you need to self-host on less hardware, want a longer advertised context, or mainly read documents and charts, where Scout matches it.

It is also worth knowing where the line stands. Behemoth never shipped, and Meta's April 2026 frontier model, Muse Spark, is proprietary. Meta has said it intends to keep releasing open-weight models, but at the time of writing Maverick remains the largest Llama you can download. Plan around it as a stable, well-understood option rather than the first step of a roadmap.