NVIDIA launches SUPER, just not the SUPER gamers wanted

NVIDIA has released Nemotron 3 Super, a 120 billion parameter Mixture-of-Experts model built for agentic AI workloads. The model uses a hybrid Mamba-Transformer design and activates 12 billion parameters during inference, which lets NVIDIA target lower serving cost than a dense model of similar size.

According to NVIDIA, Nemotron 3 Super is the first model in the Nemotron 3 series to combine LatentMoE, Multi-Token Prediction layers, and NVFP4 pretraining. The company says this stack improves accuracy and speeds up inference, with the hybrid design delivering up to 5x higher throughput and up to 2x higher accuracy than the previous Nemotron Super model. NVIDIA also lists support for context lengths of up to 1 million tokens.

Source: NVIDIA

The release is also broader than just model weights. NVIDIA is publishing post-trained, quantized, and base checkpoints in NVFP4, FP8, and BF16, alongside datasets and model recipes. The company says the wider Nemotron 3 program includes more than 10 trillion tokens of pre- and post-training data, plus reinforcement learning environments used for post-training.

Source: NVIDIA

NVIDIA positions Nemotron 3 Super for collaborative agents and high-volume workloads, which matches how it described the Super tier when the Nemotron 3 family was introduced last year. In the release material, NVIDIA points to software development, long-document analysis, and tool-calling tasks as the main target use cases for the model.

Availability is already wide for a day-one launch. NVIDIA says Nemotron 3 Super can be accessed through build.nvidia.com, Hugging Face, OpenRouter, and Perplexity, while cloud and inference partners include Google Cloud Vertex AI, Oracle Cloud Infrastructure, CoreWeave, Together AI, Baseten, Cloudflare, DeepInfra, Fireworks AI, and Modal. The model is also packaged as an NVIDIA NIM microservice for on-prem and cloud deployment.

Source: NVIDIA

Follow us on Google News Set as Preferred Source

Share