March 2026 Summaries

DeepInfra has announced its partnership with NVIDIA to launch the Nemotron 3 Super, an advanced AI model designed for complex multi-agent applications. This model, featuring a hybrid Mixture of Experts (MoE) architecture and Mamba transformer design, offers high reasoning accuracy and fast inference while operating efficiently on a single NVIDIA GPU. It boasts a 1M token context window, facilitating coherent and uninterrupted long workflows in agentic systems. The Nemotron 3 Super demonstrates significant improvements over previous models, including up to five times higher throughput and double the accuracy. It is fully open-source, allowing access to weights, training datasets, and development recipes, enabling teams to customize and fine-tune the model according to their needs. Available immediately on DeepInfra, the model supports easy deployment with minimal setup, low latency, and a usage-based pricing model, ensuring enterprise-grade security and privacy.