Llama 4 Scout and Maverick: What Meta Released and What to Check
In this article
Correction, 21 October 2026: This page previously described Llama 4 as an upcoming release in 2026. Meta released Scout and Maverick on 5 April 2025. The earlier predictions about native audio, near-zero long-context latency and planned AI Mastery benchmark testing were unsupported and have been removed.
Meta's release announcement introduced Llama 4 Scout and Maverick as downloadable, natively multimodal models using a mixture-of-experts architecture. Behemoth was described as a teacher model still in training at that announcement. A preview description is not a downloadable checkpoint.
The official model card is the reference for supported inputs, model configuration and usage restrictions. Read it alongside the exact checkpoint you intend to deploy. This article is a deployment assessment, not an AI Mastery performance benchmark.
Separate active parameters from memory requirements
Meta reported 17 billion active parameters per token for both Scout and Maverick, with 16 and 128 experts respectively. Active parameters describe the computation selected for a token; they do not mean the entire stored model occupies the memory of a dense 17B model. In its announcement, Meta distinguished Scout on one H100 with Int4 quantization from Maverick on an H100 host.
For capacity planning, start with the actual weight files and quantization. Then budget separately for the KV cache, runtime buffers, image processing and concurrent requests. A model that loads with a short prompt can still fail under your intended context length. The hardware ROI tool can help compare a proposed deployment's costs, but its output is not proof that a checkpoint fits.
Evaluate the configuration you will actually serve
Before committing to a deployment, record these decisions:
| Decision | Evidence to retain | Failure it helps diagnose |
|---|---|---|
| Checkpoint and quantization | Repository revision, filenames and runtime version | A result from a different model or conversion |
| Context and concurrency | Prompt lengths, parallel requests and peak memory | A server that loads but runs out of memory under use |
| Multimodal input | Your image types, resolutions and expected answers | Text success being mistaken for image-task reliability |
| Application behavior | Task success, invalid outputs, latency and failures | A public benchmark being substituted for your workload |
| License | The license and acceptable-use terms for the checkpoint | Download access being mistaken for unrestricted usage |
What the launch does not establish
A maximum context specification does not establish reliable retrieval throughout that context, or constant latency as it grows. Likewise, downloadable weights do not eliminate hosting, maintenance or evaluation costs. Compare the full cost of an acceptable answer, including failed attempts and operational work, with an API alternative.
Use a small set of real tasks with expected outcomes before expanding the trial. Keep failed cases rather than reporting only successful examples. That produces a deployment decision you can revisit when the model, quantization or runtime changes, without treating a launch headline as an operating guarantee.
Related Reading
Meta Muse Ember: 1M Context, 512K Output, a Third of Haiku's Cost
Meta's Muse Ember ties Claude Haiku 5.5 on the AA Intelligence Index at about a third of the cost per task, with 1M context, 512K output, and closed hosted weights.
TokenEconomizer Cuts Coding Agent Costs 18–31% Without Touching the Model
Microsoft Research and UC Berkeley's open-source plugin trims SWE-Bench spend by 18.4–30.9% across five agent–model combos, with no model changes required.
AKV Cuts Agent Serving Cost 27.7% Without Rewriting History
Tsinghua’s AKV drops and repositions KV cache instead of summarizing agent history, cutting average serving cost 27.7% on SGLang.