NVIDIA Nemotron 3 Super: 5x Faster Agentic Reasoning
NVIDIA released Nemotron 3 Super on March 11, 2026 — a 120B hybrid Mamba-Transformer MoE model with only 12B active parameters that delivers 2.2x higher throughput than GPT-OSS-120B. It scores 60.47% on SWE-Bench Verified (versus 41.90% for GPT-OSS) and maintains 91.75% accuracy on the RULER benchmark at 1 million token context. Here’s what the architecture actually does and where the caveats live.