Llama 4 brings massive context windows and open-weight ambition

When Llama 4 dropped, it arrived with the kind of spec sheet that makes open-model enthusiasts cheer and infrastructure teams immediately reach for a calculator. According to the Hugging Face release post, the launch features Maverick (roughly 400B total parameters) and Scout (roughly 109B). Both are natively multimodal Mixture-of-Experts (MoE) models with 17B active parameters, available under Meta’s community license. This is a massive moment for open weights, but it requires a heavy dose of deployment reality.

The MoE architecture is critical because the active parameter count during inference is significantly lower than the total size. This makes running these giant models less computationally impossible than a dense model of the same size. Hugging Face notes that Scout, with its 16 experts, can fit on a single server-grade GPU using on-the-fly quantization. Maverick, utilizing 128 experts, is a heavier lift. The ability to run serious, natively multimodal models locally is what moves open AI from hobbyist experiments to real-world applications.

However, open weights do not mean public domain. Meta’s licenses come with specific terms, and anyone building a business on top of Llama 4 needs to verify their compliance posture before getting too attached to the models. Furthermore, extreme context windows and multimodal inputs demand rigorous local evaluations. The true value of a major Llama release isn't Meta's benchmark claims; it's the explosion of community tooling—quantized builds, serving recipes, and fine-tunes—that follows. Llama 4 is a serious platform, provided builders test it relentlessly rather than just admiring the brochure.