NVIDIA Nemotron 3 Super scores 60.47% on SWE-Bench, topping GPT-OSS-120B by 20 points

Short summary

NVIDIA released Nemotron 3 Super at GTC March 2026, a 120B-parameter hybrid model scoring 60.47% on SWE-Bench Verified — beating GPT-OSS-120B by nearly 20 points. The model combines Mamba-2 SSM layers, transformer attention, and a novel LatentMoE architecture for 1M-token context at linear-time cost. Available free on OpenRouter and Hugging Face, it targets agentic coding workloads with 2.2x inference throughput improvement.

• •120B hybrid model scores 60.47% on SWE-Bench Verified, beating GPT-OSS-120B by ~20 points
• •Combines Mamba-2 SSM, transformer attention, and LatentMoE for 1M-token context at linear cost
• •Free on OpenRouter and Hugging Face; enterprise deployment via NVIDIA NIM containers

Generated with AI, which can make mistakes.