PlainAI AI explained plainly

Meta · 2025 · 04 · 05 · Model · ~1 min read

Meta released Llama 4 (Scout + Maverick)

Meta's first mixture-of-experts open-weights models — Scout and Maverick. The model has many sub-models inside; only the relevant ones light up for any question. Cheaper to run for what you get. Scout shipped with a 10-million-token context window, the longest of any major open-weights model at launch.

What's actually new

  • Mixture-of-experts in the open. First time the architecture shipped at this scale with downloadable weights.
  • 10-million-token context on Scout — paste a whole codebase or a stack of long PDFs.
  • Strong multilingual performance — meaningfully better than Llama 3 on non-English languages.

If you want more

Worth knowing~30s
  • Launch-day scores were cherry-picked. Independent testing showed weaker real-world performance than the headlines suggested.
  • The 10M-token context degrades quickly. Past ~1M tokens accuracy drops; past 5M it's barely usable.
  • Behemoth (the giant 2-trillion-parameter sibling) was promised but not shipped.
Who should care~20s

Companies running their own AI on their own hardware who care about cost-per-token. Teams needing long-context analysis on a budget. Researchers studying mixture-of-experts at open-weights scale.

What to do about it~20s

If you're running Llama 3.x in production, plan to test Llama 4 — costs typically drop. Validate on YOUR workload, not the launch-day numbers. New deployments should compare Llama 4 Scout for long-context work specifically.

Honest take~45s

Llama 4 was the moment open-weights AI moved into mixture-of-experts territory at the leading edge. The launch-day buzz was over-hyped and the real-world performance was slightly behind expectations, but the architectural shift mattered: it's much cheaper to run mixture-of-experts at scale than dense models, and that economics was the bigger story than any single test result. Open-source AI got a structurally cheaper future.

Other recent model updates

Sources

Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com