NVIDIA researchers have introduced two open Mixture-of-Experts models that pair Mamba and Transformer designs: Nemotron 3.5 Super, at 120 billion parameters, and the smaller 30 billion-parameter Nemotron 3.5 Lightning. The split points to two different priorities: broader multimodal and agentic capabilities in Super, and efficient text tasks in Lightning.
NVIDIA Watch analysis
What happened
The researchers describe both models as designed for reasoning, long-context understanding and agentic workloads. Super includes multimodal capabilities, continued pretraining in Indic languages and mid-training for agentic software engineering. Lightning is focused on efficient text tasks.
The researchers say the models use large-scale distillation, long-context training and native speculative decoding to improve performance and deployment efficiency. Those are the paper’s descriptions of the approach, not independent performance results.
Our top picks
- A smaller model with a distinct brief
Lightning’s 30B size and focus on efficient text tasks offer a different deployment target from the 120B Super. - Multimodal capability in Super
The larger model is described as handling more than text, broadening its intended use. - Indic-language continued pretraining
Super’s additional training is a concrete effort to extend its language coverage. - Software-engineering preparation
Super includes mid-training for agentic software engineering, aimed at work involving AI agents. - Native speculative decoding
The researchers say the technique is part of their approach to improving performance and deployment efficiency.
Why it matters
This is a model-family announcement with a useful distinction: one model is positioned for broader capabilities, while the other targets efficient text work. For developers and operators, that makes the practical question less about which model has the bigger number and more about whether either fits a specific workload and computing budget.
The paper’s description gives readers a map of the intended capabilities, but the supplied evidence does not establish comparative benchmark results or real-world performance. Those details will matter before anyone can judge how well the design translates into practice.
Our read
Two model sizes and different stated priorities are more informative than a single grand claim about doing everything. The next test is whether the technical results support the pitch, particularly for long-context and agentic tasks. “Efficient” is a useful ambition; the numbers are where it earns its keep.
What to watch
- Benchmark results and the tasks used to measure them.
- Practical memory and compute requirements for each model.
- Evidence of performance on Indic-language and agentic software-engineering tasks.
Discussion spark: For open models aimed at agentic work, should developers prioritise broad capabilities like multimodality, or smaller models that may be easier to deploy?
Sources and evidence
- NVIDIA Nemotron 3.5 Super and Lightning Hybrid Mamba-Transformer Models (30 October 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.