NVIDIA Nemotron 3 Super Outperforms GPT-OSS-120B on Coding Benchmark at GTC 2026
NVIDIA unveiled Nemotron 3 Super on March 11, 2026, at its GTC conference, releasing a 120-billion-parameter open-weight AI model targeting agentic and coding workloads. The model scored 60.47% on SWE-Bench Verified, surpassing GPT-OSS-120B's 41.90% by nearly 20 percentage points, while also delivering 2.2 times the inference throughput. It features a hybrid architecture combining Mamba-2 state space layers, standard transformer attention, and a novel LatentMoE system that activates four times more experts at the same computational cost. A 1-million-token context window and Multi-Token Prediction enable fast, long-context reasoning, with benchmarks showing up to 3x wall-clock speedup over DeepSeek-R1. The model is freely available on OpenRouter and Hugging Face, with enterprise deployment supported via NVIDIA NIM containers.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in