Meta has launched two models, Llama 4 Scout and Maverick, with a multimodal architecture and a mixture-of-experts framework to enhance various AI applications, from image processing to reasoning. Scout has 17 billion active parameters while Maverick also has the same number but leverages more experts for improved coding capabilities. These models are built from the flagship Behemoth model, boasting 288 billion parameters. While they are competitive with models like GPT-4, initial user feedback indicates performance issues, raising questions about their effectiveness.
Either they are terrible or there is something really wrong with their release/implementations. They seem bad at everything I have tried.
Both models were distilled from Meta's still-training flagship model, Llama 4 Behemoth, which has 288 billion active parameters and nearly two trillion total.
Collection
[
|
...
]