Nemotron 3.5 Super gets 51.78% on a software engineering test

The authors of Nemotron 3.5 give two open models, Super and Lightning, in a paper. Super has 120B parameters (the values that a model learns) and handles text, images and video. Lightning has 30B parameters and handles text only. The two models are for reasoning, long documents and agent work. The authors report that Super gets 51.78% on SWE-bench Pro, a test of software engineering tasks.

How it works

The two models use a mixture of experts design. Only 12B parameters in Super and 3B in Lightning work on each step. They also mix two model designs, Mamba and Transformer, and learn from large models and from reinforcement learning.

The numbers

• Super gets a score above Nemotron 3 Ultra with about 4.6 times less active parameters.
• Tests in Indian languages increase by a maximum of 9.82 points with more Indic training.
• Speed is a maximum of 4,856 output tokens a second for each GPU on GB200.

What is not known

These are the results of the authors. The abstract gives no outside test and no price. It does not compare the models with models from other makers.

Insight: Open Nemotron models include search and reasoning

NVIDIA released five open Nemotron 3.5 models to find documents on 21 October 2026. This paper gives open reasoning models in the same family. Meta opened Muse Spark on 17 October with no test results. This reads as a move to include the full agent stack in open models. The next move to monitor is an outside test. Builders can then test the models on their tasks.