Llama 4 Maverick

Llama 4 Maverick is Meta's natively multimodal language model, introduced on April 5, 2025 as part of the Llama 4 family (alongside Llama 4 Scout). It is the first Llama generation built on a Mixture-of-Experts (MoE) architecture, with alternating dense and MoE layers. It has 17B active and 400B total parameters, with 128 experts.

The model uses early fusion, combining text and image tokens into a single backbone (native multimodality). It supports up to 1M tokens of context. It was built in part through distillation from the larger Llama 4 Behemoth model (288B active, 16 experts). According to Meta, Maverick is the best-in-class multimodal model, beating GPT-4o and Gemini 2.0 Flash on a range of benchmarks, and its experimental chat version scored an ELO of 1417 on LMArena.

Weights are released under the Llama 4 Community License (open weights with restrictions, e.g. for entities with >700M monthly active users), allowing self-hosting. The model is available via Hugging Face and provider APIs (e.g. OpenRouter).