AuroraML unveils Aurora-3: a 4-trillion parameter multimodal model for enterprise developers

AI · 6 min read

AuroraML today announced Aurora-3, a 4-trillion parameter multimodal foundation model designed to handle text, audio, images, and structured data in low-latency production environments. The model introduces a new sparse attention layer and dynamic routing that AuroraML says cuts inference costs by up to 60% compared with dense alternatives while enabling single-model support for cross-modal tasks.

Alongside the product launch, AuroraML disclosed a $120 million Series B led by Sentinel Capital with participation from enterprise cloud funds. The proceeds will fund global deployment of AuroraLab, the startup’s managed inference edge that bundles model hosting, MLOps, and compliance tooling aimed at regulated industries.

AuroraML positions Aurora-3 for real-world adoption with prebuilt adapters for retrieval-augmented generation, document understanding, and voice agents. The company also released a developer SDK and a tuned instruction set for customer fine-tuning, promising shorter iteration cycles and transparent cost controls for enterprise customers.