Llama 4 Maverick
By Meta
Llama 4 Maverick is a large, multimodal open-weight model in Meta's Llama 4 family, built on a mixture-of-experts architecture with a bigger active-parameter budget than the companion Llama 4 Scout model, aimed at delivering stronger…
Definition
Llama 4 Maverick is a large, multimodal open-weight model in Meta's Llama 4 family, built on a mixture-of-experts architecture with a bigger active-parameter budget than the companion Llama 4 Scout model, aimed at delivering stronger general and multimodal reasoning capability. It shares Llama 4 Scout's mixture-of-experts design but activates a larger number of parameters per token, trading higher compute cost for stronger general and multimodal reasoning quality across text and image inputs.
Overview
Llama 4 Maverick is the larger and more capable sibling to Llama 4 Scout within Meta's fourth-generation Llama lineup, sharing the same mixture-of-experts architectural approach but with a bigger pool of experts and a larger active-parameter footprint per token, aimed at pushing general reasoning, coding, and multimodal understanding closer to leading proprietary models. Maverick uses the same expert-routing mechanism as Scout, but with a larger pool of experts and a bigger active-parameter budget engaged per token, so more of the model's total capacity is brought to bear on each request at the cost of higher compute per token than Scout's more economical routing. Like Scout, Maverick accepts both text and image inputs, continuing Meta's multimodal direction started in Llama 3.2, and it is designed to handle tasks that combine visual and textual reasoning, such as interpreting charts, diagrams, and mixed-media documents, alongside standard text-based conversation and problem solving. Its multimodal input pipeline mirrors Scout's approach of pairing an image encoder with the language model backbone, but the larger active-parameter budget behind it is intended to translate into stronger performance on harder combined visual-and-text reasoning tasks rather than introducing a fundamentally different multimodal mechanism. The mixture-of-experts design lets Maverick maintain a very large total parameter count while activating only a subset of experts for each token, which keeps inference more efficient than a fully dense model of equivalent total size would be, though Maverick's larger active-parameter budget compared to Scout means it generally requires more compute per request in exchange for stronger output quality. Where Scout is tuned toward long-context efficiency, Maverick is tuned toward raw output quality, so the practical choice between the two Llama 4 siblings comes down to whether an application is more constrained by document length and serving cost or by the ceiling of reasoning and multimodal capability it needs. Maverick was positioned by Meta as competitive with leading models from other labs on a range of benchmarks at the time of release, and being open-weight, it can be downloaded, fine-tuned, and deployed on private infrastructure, appealing to organizations that need both strong capability and control over data handling. Organizations building high-stakes multimodal assistants, technical document analysis tools, or coding systems that need the strongest available open-weight quality have deployed Maverick in place of a closed flagship API where self-hosting and data control were priorities, accepting the higher compute cost that comes with its larger active-parameter footprint. Within the Llama 4 family, the choice between Scout and Maverick generally comes down to a trade-off between context length and efficiency versus raw general and multimodal capability, letting developers pick the variant that best matches their latency, cost, and quality requirements. Running Maverick well requires more substantial mixture-of-experts-aware infrastructure than Scout, and because it activates more parameters per token, its per-request cost and latency are higher, so teams that do not need its extra capability on most requests often route easier queries to Scout or a smaller model and reserve Maverick for harder cases.
Key Features
- Mixture-of-experts architecture with a larger active-parameter budget than Scout
- Multimodal support for combined text and image reasoning
- Strong general benchmark performance relative to other open models
- Open weights available for self-hosted deployment and fine-tuning
- Designed to handle mixed visual and textual document understanding
- Positioned as the more capability-focused sibling to Llama 4 Scout