NVIDIA said on March 12 it released Nemotron 3 Super, a 120B-parameter open-weights model with 12B active parameters built for agentic AI, pushing the company deeper into enterprise AI workflows beyond chips. For more on Nvidia’s broader AI momentum, see this roundup ahead of GTC 2026.
The company describes the model as a hybrid Mamba-Transformer Mixture-of-Experts architecture with a 1M-token context window, and says it is publishing the weights, datasets, and training recipes; VentureBeat reported the weights are posted publicly.
On NVIDIA’s own benchmarks, Nemotron 3 Super delivers up to 2.2x and 7.5x higher inference throughput than GPT-OSS-120B and Qwen3.5-122B at 8k input/16k output, plus up to 5x throughput and 2x accuracy versus the previous Nemotron Super—results that still await independent validation.
The Register notes NVIDIA is positioning a Nemotron family (Nano/Super/Ultra) as open models, and if third-party benchmarks confirm the vendor claims, the combination of open weights and 1M-token context could shift enterprise agent adoption; the next step is clearer licensing and external evaluation.
