Meta's Llama 4 Scout and Maverick are the open-weight bookends of April 2026: Scout pushes context length to 10 million tokens, Maverick pushes parameter scale to 400 billion via mixture-of-experts. Together they reset what is possible inside the open-weight ecosystem and, more importantly for brand visibility, reset what runs inside the thousands of self-hosted enterprise apps that prefer Meta's licensing terms over Anthropic's or OpenAI's.
Scout: the 10M context model that actually runs
Scout is the surprise. A 17B-active-parameter model with effective 10M-token context that runs on a single H100 with the right quantization. The technical achievement is interleaved attention with a chunk-aware routing layer that keeps memory footprint sublinear. The brand visibility consequence is that Scout becomes the default for enterprise document-search and competitive-intelligence workflows where teams want to keep all their data on-prem and have not been able to do so until now.
Practically: when a private-equity analyst dumps a 1,200-page deal room into Scout and asks "list every vendor mentioned, with their stated capabilities," every brand that appears in the deal room gets normalized into a structured list. Coverage in obscure sources (board decks, due-diligence reports, regulatory filings) suddenly counts toward brand surface area in a way it did not before.
Maverick: the 400B MoE for serving
Maverick is 400B parameters total, with 17B active per token (MoE-128 routing). It posts frontier-comparable benchmarks (MMLU 88.4%, MATH 91.2%, HumanEval 89.5%) at a per-token serving cost competitive with Mistral Large 2 and Qwen 3.6-Max. It is the model that gets fine-tuned by enterprises that want a frontier-class model under a permissive license they can audit.
Brand visibility consequences of the Llama 4 family
Three things shift. First, the training corpus rebalances toward open-source code, Meta's own data partnerships, and a heavily filtered Common Crawl. Brands strong in the open-source code commons gain. Brands that depend on news-syndicated coverage may drop. Second, the licensing model means Llama 4 will get embedded in self-hosted RAG and agent stacks at large enterprises who specifically chose it over closed alternatives. You cannot directly monitor those deployments. Third, fine-tuning on Llama 4 is now economically viable for mid-market companies, which means private domain models will start showing up inside categories where general-purpose models used to dominate. Brand recall in those private models is the result of who ingested what during fine-tuning.
What to test this week
Run a brand-recall test on Llama 4 Maverick via Together AI or Replicate. Run the same on Scout if you have access. Compare against your closed-model baseline. If Llama 4 gives you significantly thinner coverage than GPT-5.5 or Claude 4.7, your gap is in open-source code repositories, GitHub discussions, technical documentation, and well-cited Wikipedia presence.