NVIDIA's Nemotron 3 Super Beats GPT-OSS-120B on Coding Benchmarks
NVIDIA's open Nemotron 3 Super model beats GPT-OSS-120B by 20 points on SWE-Bench and delivers 2.2x faster inference throughput.
At GTC 2026, NVIDIA released Nemotron 3 Super, a 120-billion-parameter open-weight hybrid model built for agentic coding workloads. On SWE-Bench Verified — one of the toughest real-world coding benchmarks — it scores 60.47%, nearly 20 points ahead of GPT-OSS-120B's 41.90%, while delivering 2.2x higher inference throughput.
The architecture combines Mamba-2 state-space layers for linear-time long-context processing, standard transformer layers for precise reasoning, and a novel LatentMoE routing system that activates four times more experts at the same compute cost. Multi-Token Prediction adds further speed gains through more efficient speculative decoding. The payoff shows up at scale: a 91.75% RULER score at a full 1-million-token context (versus 22.30% for GPT-OSS-120B) means the model actually retains and reasons over entire codebases, not just accepts long inputs nominally.
For engineering teams, the throughput and accuracy gains translate directly into lower per-request infrastructure costs — NVIDIA cites potential savings of 50-80% versus comparable open models. Nemotron 3 Super is available free via OpenRouter, self-hostable through Hugging Face under NVIDIA's Open Model License, and deployable at enterprise scale via NVIDIA NIM. Tools like CodeRabbit, Factory, and Greptile are already integrating it for code review and repository-scale analysis.