Llama 4 brings native multimodality to a new open-weight generation

Scout and Maverick arrive

Meta introduced Llama 4 on April 5, 2025, releasing Scout and Maverick and discussing a larger Behemoth model that was still in development. The released models used mixture-of-experts designs and native multimodal capabilities.

An announced research model and downloadable production candidates are different categories. A careful comparison identifies which artifact is actually being tested.

A context headline needs an application test

Large advertised context capacity can make repository analysis or document collections seem straightforward. In practice, input preparation, memory use, and the ability to retrieve a relevant detail still matter. A model that accepts a large input is not guaranteed to resolve every contradiction inside it.

Image understanding also needs to be checked on the types of assets the application handles, from clean product photos to dense annotated layouts.

Record the configuration

Save the checkpoint revision, serving framework, quantization, and context settings with each evaluation. Compare a concise curated input with a much larger unfiltered one. That test helps reveal whether extra context improves the result or mainly increases cost and complexity. Review the model's license and distribution conditions before treating an experiment as a deployable service.