NVIDIA details Nemotron 3.5 Super and Lightning, open hybrid Mamba MoE models for agents

What NVIDIA published

NVIDIA submitted its technical report on Nemotron 3.5 Super and Nemotron 3.5 Lightning to arXiv on 28 October 2026, with 557 listed authors. Both are open mixture-of-experts models built on a hybrid of Mamba-2 and Transformer layers and aimed at reasoning, long context and agent workloads. Super has 120 billion total and 12 billion active parameters and reads text, images and video. Lightning has 30 billion total and 3 billion active parameters and handles text only.

The weights were already public on Hugging Face: the Super VL model card gives 15 October 2026 as its release date and the Lightning model card gives 11 August 2026. Both cards list a context of up to 1M tokens and the OpenMDW 1.1 license, which allows commercial use. The report adds the training recipe, the evaluation setup and the efficiency numbers behind those releases.

How the two models were trained

NVIDIA started Super from Nemotron 3 Super Base and Lightning from Nemotron 3 Nano Base, then distilled each for 1 trillion tokens from the next larger Nemotron 3 model. Super then received 100 billion tokens of Indic continued pretraining, 67 billion tokens of software-engineering mid-training at 256K context and a short 1M-context recovery stage. NVIDIA says the Indic stage lifts Indian-language benchmarks by up to 9.82 percentage points and that it generated about 2.83 trillion tokens of Indic text across 22 languages.

Post-training used supervised fine-tuning and reinforcement learning, and for Super a multi-teacher on-policy distillation step that turns a text-only base into a vision-language agent. Both models carry a multi-token prediction head for speculative decoding. NVIDIA reports that post-training boosting raises the average draft acceptance length by up to 21.8% under probabilistic sampling, and that the NVFP4 build keeps a median 99.62% of BF16 accuracy across 36 benchmarks.

What the numbers say

NVIDIA's headline claim is 51.78% on SWE-bench Pro for Super, above the 49.66% its model card lists for Nemotron 3 Ultra, which uses about 4.6 times more active parameters. The same table shows open rivals ahead on several agent benchmarks: Qwen3.8 Flash-Next scores 60.37% on SWE-bench Pro and 85.80% on Terminal-Bench 2.1, against 60.67% for Super. On a single GB200 GPU, NVIDIA reports up to 4,856 output tokens per second on a 20-turn agent workload while keeping at least 50 tokens per second per user. These are NVIDIA's own measurements, run through its NeMo Gym harness.

Why it matters and what to watch

The report makes Nemotron 3.5 one of the more fully documented open model families: weights, drafters, data and recipes are published, not only the checkpoints. That matters for companies that want to retrain or audit a model rather than call an API. Super VL is also the parent of the 24GB Nemotron 3.5 Bolt local model. NVIDIA says it will release a broader Indic synthetic corpus, a companion report and, with AI4Bharat, a human-verified Indic evaluation suite; dates are not stated. Independent benchmark runs will show how the efficiency claims hold outside NVIDIA's setup. More open-weight releases are tracked under Open Source AI and in the model release tracker.

Questions people ask

What is the difference between Nemotron 3.5 Super and Lightning?

Super is a 120B-parameter model with 12B active parameters that reads text, images and video. Lightning is a 30B text-only model with 3B active parameters that fits on a single 80GB GPU.

Are Nemotron 3.5 Super and Lightning open source?

The weights are open under the OpenMDW License Agreement 1.1, which the model cards say allows commercial use. NVIDIA also releases pretraining and post-training datasets and training recipes.

How well does Nemotron 3.5 Super score on coding benchmarks?

NVIDIA reports 51.78% on SWE-bench Pro and 60.67% on Terminal-Bench 2.1. Its own comparison table shows some open models, such as Qwen3.8 Flash-Next, scoring higher on both.

Sources

1. Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI (NVIDIA (arXiv), 28 Oct 2026)
2. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 model card (NVIDIA on Hugging Face, 15 Oct 2026)
3. NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 model card (NVIDIA on Hugging Face, 11 Aug 2026)