Meta Llama 4 family landed with a headline number that is hard to ignore: Scout, the smaller variant, offers a ten-million-token context window. That is the largest of any model, open or closed. It is also a feature most people quoting it will never actually use.
The numbers
Llama 4 comes in tiers. Scout is 109 billion total parameters across 16 experts, with that ten-million-token window. Maverick is bigger, 400 billion total. Both ship under the Llama 4 Community License, which is open enough for most uses and not quite OSI-open, a distinction Meta keeps hoping nobody presses on.
What ten million tokens really means
Ten million tokens is roughly a shelf of books, or an entire mid-sized codebase, in a single prompt. Sounds incredible. In practice, two things bite. First, quality degrades across very long contexts. Models get lost in the middle, and the reliable window is smaller than the maximum advertised. Second, feeding millions of tokens is slow and expensive, so even when it works, you often would not want to.
A giant context window is like a giant truck bed. Nice to have once a year, and mostly you are driving to the shop for milk.
For the overwhelming majority of real tasks, a few thousand tokens of the right context beats ten million tokens of everything. The skill that matters is choosing what to put in the window, not bragging about how big it is.
Where Llama 4 actually earns it
- Running it yourself. Llama is still the most boringly reliable open family to deploy, with a huge ecosystem.
- Cost and control. No per-token bill, no vendor who can switch you off, full visibility. That is the real Llama pitch.
- The occasional genuine haystack. Once in a while you really do need to reason across a massive document set. On those days, Scout is a gift.
A good, useful release. Just do not choose it for a spec you will use twice a year. Choose it because you want to own your stack.