NVIDIA announced Nemotron 3.5 Super and Nemotron 3.5 Lightning on October 14, 2026, and says it is releasing model checkpoints and supporting materials. Super is a 120-billion-parameter model that uses 12 billion parameters per token; Lightning is a text-only model with 30 billion total parameters and 3 billion active per token. Both use a hybrid design that combines Mamba and Transformer components with a Mixture-of-Experts setup, which activates only part of the model for each token.
The models are aimed at long documents and agent-style tasks, where software can use tools and take multiple steps. Super can work with text, images, video and interactive tools. NVIDIA says both models support context lengths of up to 1 million tokens. Its release includes BF16 and compressed NVFP4 checkpoints, plus training and evaluation recipes, datasets and speculative-decoding tools. NVIDIA’s announcement does not specify a price or license terms.
What NVIDIA says Super can do
NVIDIA reports that post-trained Super scored 51.78% on SWE-bench Pro, a software-engineering benchmark, beating Nemotron 3 Ultra while using about 4.6 times fewer active parameters. The company also says agent-focused training raised DeepSWE pass@1 from 6.4% to 10.4%. These are NVIDIA’s reported evaluations; the supplied materials do not include independent validation.
The company describes a substantial training process behind Super. It used continued training on data covering 22 Indian languages and 34 language-script combinations, which NVIDIA says improved five evaluated Indic-language benchmarks by as much as 9.82 percentage points. Further training targeted long-context tasks and software engineering. The company says a final training stage restored Super’s RULER-1M score from 33.5 to 81.0 after an earlier stage lowered it.
Performance depends on the setup
NVIDIA also tested running several agents in parallel and having them check their own answers. In the company’s evaluated setups, this raised Terminal-Bench 2.1 from 62.7% to 71.9%, DeepSWE from 24.6% to 34.5%, and BrowseComp from 71.5% to 79.8%. These figures describe that multi-agent approach, not a guaranteed result for every deployment.
For speed, NVIDIA reports up to 4,856 output tokens per second on one GB200 GPU while maintaining at least 50 tokens per second per user in its tested 20-turn agent workload. Its NVFP4 version recovered a median 99.62% of BF16 accuracy across 36 benchmarks, according to the company. Those results depend on NVIDIA hardware and the test setup; they do not establish how the models will perform on other systems.
Why it matters
For researchers and developers, the released checkpoints and recipes offer a way to try the models or build on them without training from scratch. NVIDIA’s benchmark results suggest Super may suit software-engineering agents, but independent results, licensing terms and the hardware needed for a practical deployment remain unclear from the supplied details.
