At long last, Meta has released the new Llama 4 family of models, and the announcement is here. The models in this family come in three classes:

1. Llama 4 Scout - 17 billion parameters, 109B total parameters, fits in an H100 GPU
2. Llama 4 Maverick - 17B active parameters, 400B total parameters
3. Llama 4 Behemoth - 288B active parameters, 2T total parameters, not yet released

These models are particularly interesting because they’re multimodal and their architecture is based on the mixture of experts (MoE) approach. This splits the model into parts (experts), feeds input tokens to each expert, and combines their outputs to form the final result.

The Llama 4 Behemoth model looks incredible, but it’s not available just yet. So I’ll download Llama 4 Maverick and see if it’s as amazing as Meta claims it to be.