Meta releases Llama 4.1: Scout, Maverick, and Behemoth MoE models under Apache 2.0

Think of Mixture of Experts (MoE) like a large office where every incoming request gets routed to the most suitable specialist, rather than everyone working on every task at once. That is the core idea behind Meta's new Llama 4.1 family: instead of activating the entire neural network for every response, only a relevant subset of "experts" is engaged, making the models much more efficient.

Meta released three distinct versions. Scout (17B parameters, 16 experts) is designed to run on edge devices — meaning regular computers or compact hardware — without needing cloud infrastructure. Maverick (17B parameters, 128 experts) is the mid-tier option but with a remarkable capability: it can hold up to 10 million tokens in its context window, roughly equivalent to 7,500 pages of text, or an entire company codebase. Behemoth (288B parameters, 128 experts) is the frontier model, targeting research labs and high-end data centers.

All three models understand text, images, and audio natively — no separate add-on components needed.

The most important aspect of this release is the licensing: all models ship under Apache 2.0. This means anyone — companies, researchers, independent developers — can use, modify, and redistribute them freely, including for commercial products, with no royalties. This is a landmark moment for open source AI, significantly narrowing the gap between openly available models and the proprietary systems offered by major cloud providers.