What the family record covers
Meta's Llama documentation currently identifies Llama 4 Scout and Llama 4 Maverick as multimodal models for image and text understanding. The family can be obtained directly or through multiple model and cloud partners.
Downloadable weights create deployment choices that hosted-only comparisons miss. Hardware, quantization, serving software, prompt format, safety layers, and operator updates become part of the evaluated system.
Documented access paths
- Model access directly from Meta or named distribution partners
- Self-managed or third-party hosted inference under the applicable model license
- Separate Llama Guard resources for safety-layer evaluation
Capabilities to verify
- Image-and-text behavior for the selected checkpoint and prompt format
- Throughput, memory, and latency on the actual serving stack
- Quantization impact on task quality and structured output
- Safety controls, monitoring, and upgrade ownership in the chosen deployment
Decision questions
- Will the organization self-host or use a named inference provider?
- Which checkpoint, license terms, quantization, and serving version define the pilot?
- Who owns safety-layer configuration, patching, and model replacement?
Evaluation cautions
- The same checkpoint can behave differently across serving and quantization choices.
- Open or downloadable access does not remove license, security, or governance obligations.
- Provider benchmark claims need independent replication on the intended stack.
Related decisions
Move from a family name to the product, provider, workflow, comparison, or protocol that defines the real choice.
Sources checked
Open the original pages before relying on a time-sensitive product decision.