Nvidia has launched its most advanced AI model yet, Nemotron Super 3. Designed for large-scale agentic AI systems, this new model combines sophisticated reasoning with rapid processing.
Nemotron Super 3, a 120 billion-parameter open model, delivers up to five times higher throughput and double the accuracy of its predecessor. This advancement addresses critical challenges in agentic AI, including managing massive data generation and the "thinking tax" associated with complex reasoning.
With a one million token context window, Nemotron Super 3 maintains full workflow state, preventing "goal drift." Crucially, only 12 billion parameters are active during inference, significantly reducing computational cost and increasing speed.
The model runs on Nvidia's Blackwell GPUs, offering up to four times faster inference compared to previous platforms. Nemotron Super 3 is readily available for download and is being integrated by AI search engines like Perplexity, coding applications, and life science organizations for advanced data analysis.
Major companies across telecommunications, cybersecurity, and semiconductor design, including Amdocs and Palantir, are adopting Nemotron Super 3 to automate complex workflows. Dell and HPE will also provide access through their agent hubs.
This release precedes Nvidia's annual GTC conference, where new GPU platforms are anticipated.