Llama 4 Scout is easy to review badly. You quote the ten-million-token context window, call it revolutionary, and move on. Having actually run it, the honest review is more interesting, and more useful: Scout is a very good open model whose best feature is not the one on the box.
The headline feature, tested
Scout carries 109 billion parameters across 16 experts and, yes, a ten-million-token context window. In practice, the usable window, where answers stay reliable, is smaller than the maximum, the same "lost in the middle" problem every long-context model has. Feeding it millions of tokens is also slow and costs real compute. It works. You just will not want to do it often.
What actually makes it good
Freedom. Scout runs on your hardware, under an open license, with no per-token bill and no vendor who can switch you off. For a huge range of real tasks it is plenty capable, and you own the whole thing. That is the feature that changes how you build, not the context number.
The best thing about Scout is not that it can read ten million tokens. It is that nobody can take it away from you.
Where it falls short
Against the absolute frontier, Scout is a step behind on the hardest reasoning and code. If your work needs the last few percent of capability, this is not the model. For the ordinary 90 percent, run on your own terms, it is excellent value.
The verdict
- Best for: self-hosting, privacy, cost control, and a mature ecosystem that just works.
- Skip if: you need frontier-level reasoning, or you chose it purely for the ten-million-token spec.
- Reality check: use a few thousand tokens of the right context, not ten million tokens of everything.
Scout is a genuinely good open model. Pick it because you want to own your stack. Treat the context window as a party trick you will occasionally, delightfully, need.