NVIDIA open-sources the 120B Nemotron 3 Super intelligent agent model, with throughput increased by up to 5 times

Gate News, March 12 — NVIDIA announced the open-source large language model Nemotron 3 Super, designed for multi-agent application scenarios. The model has a total of 120 billion parameters, utilizing a hybrid Mamba-Transformer MoE architecture. During inference, only 12 billion parameters are activated per token. Its core technology, “Latent MoE,” compresses token embeddings into a low-rank latent space before routing to expert networks, enabling four experts to be activated simultaneously at the cost of a single expert’s computation. This results in up to a fivefold increase in inference throughput compared to the previous Nemotron Super. The model natively supports a context window of 1 million tokens, suitable for autonomous agents that need to maintain long-term workflow states. In the PinchBench benchmark for evaluating agent workloads, Nemotron 3 Super scored 85.6%, the highest among similar open-source models. NVIDIA also open-sourced over 10 trillion tokens of training data, 15 reinforcement learning training environments, and evaluation schemes, under the NVIDIA Nemotron Open Model License. The model is now available on platforms such as Hugging Face, build.nvidia.com, Perplexity, and OpenRouter, and supports deployment via Google Cloud, Oracle, AWS Bedrock, Azure, and other cloud services. Companies like Perplexity, CodeRabbit, Cadence, Dassault Systèmes, and Siemens have already adopted it.