In addition to having the most powerful AI hardware, NVIDIA is also actively involved in open source large models. It launched the Nemotron 3 series at the end of last year, and today it released a new generation of Nemotron 3 Super. Nemotron 3 Super is said to be specially built for AI agents. It is not large in scale, with 120 billion parameters and 12 billion activations. It is claimed that the inference speed is increased by 3 times and the throughput is increased by 5 times.
This model supports 1M context and solves the problems of multi-agent context explosion and target deviation.
The innovative Mamba-MoE hybrid architecture solves the performance bottleneck in multi-Agent collaboration.
It is worth mentioning that in addition to supporting BF16 and FP8, Nemotron 3 SuperIt can also support NVFP4 training on its own Blackwell and later platforms.
As for the performance of the model, it is natural to have a wave of rankings. NVIDIA said that Nemotron 3 Super has topped the list of Artificial Analysis in terms of efficiency and openness. It leads the accuracy rate among models of the same size and drives the NVIDIA AI-Q research agent to rank first in the DeepResearch Bench and DeepResearch Bench II rankings.
In terms of ecology, it has been integrated by Perplexity, Palantir, Siemens, Dell, HPE and other companies, and can be used on AWS, Azure, Google Cloud and other platforms.
For ordinary people, Nemotron 3 Super also has a special meaning, that is, it is a perfect match for the currently popular crayfish OepnClaw.It achieved a high score of 85.6% in the OpenClaw task success rate.The performance is close to Claude Opus 4.6 and GPT-5.4, but it is open source and free, and the cost is much lower.
Open source address: https://huggingface.co/collections/nvidia/nvidia-nemotron-v3
