AI Radar: Meta Llama 4 — Multimodal, Open, With Fine Print

No open ecosystem is more widely deployed than Llama. With its fourth generation, Meta moves to mixture-of-experts and makes the models multimodal from the ground up — a clear leap, but not without fine print.

Open competition has passed Llama 4 by: Qwen3.6 27B reaches more than double the index score. (Quelle: Artificial Analysis — model leaderboard)

The Herd at a Glance

Llama 4 Scout: 17B active parameters across 16 experts, fits on a single H100, and ships with an exceptionally large context window.
Llama 4 Maverick: also 17B active parameters but 128 experts — the all-rounder that keeps pace with much larger proprietary models.
Llama 4 Behemoth: at 288B active parameters, Meta's strongest model yet, still in training and intended as a teacher for the smaller variants.
Natively multimodal: text and image are processed jointly, not bolted on afterward.

Active Parameters, Total Size, and What Each Means

With Llama 4, a look at the MoE math pays off, because it decides the hardware. The 17B active parameters of Scout and Maverick say how much compute actually falls per token — that is the figure that drives inference speed and token cost. The total parameter count sits markedly higher, because each model holds many experts in reserve and activates only the best-fitting ones per token. That total size, in turn, determines how much VRAM is needed simply to load the weights. In short: the active figure drives runtime, the total figure drives memory.

Concretely: Scout brings a 17B-active / 109B-total profile across 16 experts, so it fits on a single H100. Maverick stays at 17B active but grows to roughly 400B parameters spread across 128 experts — and needs correspondingly more hardware. Behemoth (around 288B active at an estimated two trillion total) is still in training and is not intended for productive SME deployment; it serves as a teacher for the smaller models.

The Caveat

Llama is available in the open, but not in the classic open-source sense: it carries a community license with restrictions — for very large providers and in its acceptable-use policy. For European companies, there is the added point that Meta has at times attached regional conditions to the multimodal features. So anyone deploying Llama 4 in production should vet the license as carefully as the model.

What the License Says

The Meta Llama license is more permissive than many proprietary offers, but it is plainly not a pure Apache or MIT license. Two clauses deserve particular attention. First, the 700-million monthly active user threshold: anyone exceeding it needs a separate agreement with Meta — irrelevant for the SME sector, a real hook for very large platforms. Second, the acceptable-use policy, which excludes certain applications and governs your own responsibility for operations. Anyone deploying Llama 4 commercially should have read both clauses and built them into internal guidelines — then self-hosting is unproblematic.

Our Take

The fact that Scout fits on a single H100 makes serious open AI accessible to SMEs — that is the good news. The license is the price: affordable, but required reading. We judge models by task — and with Llama, the fine print is part of the task.