They look like a small/large ladder, but they're built for different machines. Why Scout is the home model, why the smaller one has the bigger context, and the rare case for Maverick.
People treat Scout and Maverick like a good/better/best ladder where you'd obviously want the bigger one if you could afford it. For running at home, that framing is wrong, and it'll waste your money. These two Llama 4 models aren't small-and-large versions of the same thing you'd choose between on budget — they're aimed at completely different machines. Let me make the local decision simple: for almost everyone reading this, the answer is Scout, and I'll explain exactly when it isn't.
Scout and Maverick aren't budget tiers — they're built for different machines. Illustration by Aliteq. · Illustration by Aliteq / generated with Higgsfield
The difference that matters for local
Both models activate the same 17B parameters per token, so per-token compute is similar. What separates them is total size — the amount you must hold in memory. Scout's 109B total is a stretch-but-doable home target once quantized; Maverick's 400B total is roughly four times the memory, which pushes it out of single-machine home reach and into multi-GPU servers or the cloud. If you've internalised the MoE rule from the hub — you load the total, not the active count — this is just that rule deciding the whole matchup.
Scout vs Maverick, for a local decision
Total / active params
Scout
109B / 17B
Maverick
400B / 17B
Experts
Scout
16
Maverick
128
Context window
Scout
Up to 10M tokens
Maverick
Up to 1M tokens
Home-runnable (quantized)
Scout
Yes — 32GB GPU / unified memory
Maverick
No — data-center / multi-GPU
Where to run it
Scout
Your machine
Maverick
Cloud / API
Scout
Maverick
Total / active params
109B / 17B
400B / 17B
Experts
16
128
Context window
Up to 10M tokens
Up to 1M tokens
Home-runnable (quantized)
Yes — 32GB GPU / unified memory
No — data-center / multi-GPU
Where to run it
Your machine
Cloud / API
When would you ever pick Maverick?
Rarely, and basically never locally. Maverick's larger expert pool can give it an edge on some general-quality tasks, so if you're serving at scale in the cloud and need that last bit of capability from a Llama model specifically, it's there. But for an individual running at home, the honest path is Scout on your own hardware, and — if Scout's quality isn't cutting it — a jump to a better-reviewed model like Qwen3 or DeepSeek rather than to Maverick. Bigger isn't the fix when the smaller sibling already matches it on active compute and beats it on context.
Quick answers
Should I run Scout or Maverick locally?
Scout, almost certainly. It's the one that fits a home machine when quantized (32GB GPU or unified memory), and it actually has the larger context window. Maverick's 400B total is a data-center model — run it via the cloud if you ever genuinely need it.
Why is Maverick so much harder to run if it also uses 17B active?
Because mixture-of-experts models load all parameters, not just the active ones. Maverick's 400B total is about four times Scout's 109B, so it needs roughly four times the memory to hold — even though per-token compute is similar.
Which has the bigger context window?
Scout — up to 10M tokens, versus Maverick's 1M. If long context is your goal, the smaller, home-runnable model is also the better one for it.
Is Maverick worth it for quality?
For an at-home user, generally no. If Scout's quality falls short for you, a better-reviewed model like Qwen3 or DeepSeek is a smarter move than the much harder-to-run Maverick.
