Meta releases Llama 4 family with native multimodal support

Meta released the Llama 4 family, its first natively multimodal open-weight foundation models, supporting text, image, audio, and video understanding within a unified architecture. The release includes three size tiers — Scout (17B active parameters), Maverick (17B active with more experts), and Behemoth (a research-scale model not released for general use) — and continues the Llama programme's open licensing model that allows commercial deployment without per-token API fees.

**What native multimodal means for enterprise deployment.** Prior Llama releases required separate models for image understanding: a text model plus a vision encoder bolt-on, typically CLIP or a fine-tuned variant. Llama 4's native multimodal architecture processes text, images, and audio through the same transformer stack — which simplifies deployment architecture and allows a single inference endpoint to handle mixed-modality inputs without routing logic between models. For APAC enterprises processing documents that combine text, tables, charts, and stamps (common in financial, legal, and manufacturing contexts), this is a meaningful practical improvement.

**Open-weight implications for APAC regulated sectors.** Llama 4's commercial licence allows deployment on private infrastructure without data egress to Meta or any cloud provider. This directly addresses the data residency requirements that have made cloud-hosted multimodal models (GPT-4o Vision, Claude Sonnet Vision) difficult to deploy in healthcare, government, and financial services contexts where customer data cannot leave the jurisdiction. For Hong Kong, Singapore, Japan, and South Korean regulated sectors, self-hosted Llama 4 Scout or Maverick is now a credible option for document intelligence workloads.

**Performance relative to closed models.** At the Scout tier (17B active parameters), Llama 4 benchmarks below GPT-4o and Claude 3.7 Sonnet on complex reasoning and instruction-following tasks but performs comparably on structured extraction, classification, and document summarisation. For most production document processing workloads — the primary use case in APAC mid-market AI deployments — this performance tier is sufficient.

**AIMenta's editorial read.** Llama 4's native multimodality closes the capability gap that previously made open-weight models a poor choice for document-heavy APAC workflows. Enterprises with data residency requirements and existing inference infrastructure should run a formal evaluation against their specific document types before making a platform decision.