NVIDIA published technical details and benchmarks for the Nemotron 3.5 Super and Lightning models
NVIDIA has published a technical description and benchmark results for the Nemotron 3.5 Super (120B/12B active) and Lightning (30B/3B active) models. The Lightning model itself was released by the company as open source under the OpenMDW 1.1 license back on 11 August, together with the NeMo Switchyard routing tool.
NVIDIA published a technical article describing the architecture and results of the Nemotron 3.5 Super and Nemotron 3.5 Lightning models. According to the document, both are hybrid Mixture-of-Experts models combining Mamba and Transformer architectures. Super has 120 billion parameters with 12 billion active and is multimodal (text, images, video, tool interaction), while Lightning has 30 billion parameters with 3 billion active and is intended for text. Both models use long context during training and native speculative decoding; the post-training pipeline includes supervised fine-tuning, RLVR, and RLHF, with Super additionally undergoing special training of teacher models and MOPD distillation. According to NVIDIA, the Super model achieves 51.78% on the SWE-bench Pro benchmark, surpassing the previous Nemotron 3 Ultra model with approximately 4.6 times fewer active parameters; NVFP4 quantization achieves 99.62% accuracy retention compared to BF16, and inference optimizations enable up to 4 856 output tokens per second per GPU on the GB200 platform.
NVIDIA itself released the Nemotron 3.5 Lightning model back on 11 August as an open model with fully published weights, code, training recipes, and data under the Linux Foundation OpenMDW 1.1 license. It is part of the Nemotron 3 family, and in terms of size (30 billion parameters) it fits between the smaller Nano model (8 billion) and the largest Ultra (550 billion). At the same time, the company also released the NeMo Switchyard library for intelligent routing of queries among open, proprietary, and custom models in agentic applications; according to the company, the models are intended for specialized tasks such as code review, tool use, monitoring security alerts, or answering billing questions, with the ability to run locally on RTX PC, DGX Spark, or Jetson devices.
The release came less than a month after NVIDIA CEO Jensen Huang's public defense of open models amid the debate over Chinese open models (Alibaba Qwen, Moonshot AI), and a day after Meta released a new open model. Analysts cited in the sources note that open models help NVIDIA sell its own hardware and infrastructure, and that publishing training data is rather the exception among vendors.
Why it matters
The technical description provides concrete measurable figures (performance on SWE-bench Pro, inference speed, accuracy under NVFP4 quantization), which companies can use to assess whether deploying open models is worthwhile instead of paid APIs with unclear token costs. The fully published weights, code, and training data also enable local deployment with control over data, which is relevant for organizations with data sovereignty or security requirements.
What was added since the original report
Verified updates
-
Lightning is a MoE hybrid Mamba-Transformer model with 3B active parameters out of 30B total; Super is a 120B model with 12B active parameters, multimodal, intended for code engineering; Both models include long-context training and native speculative decoding; The post-training pipeline uses RLVR, RLHF, and for Super, multi-teacher on-policy distillation (MOPD); Super achieves 51.78% on SWE-bench Pro and has 4.6× fewer active parameters than Nemotron 3 Ultra
- Lightning is a MoE hybrid Mamba-Transformer model with 3B active parameters out of 30B total
- Super is a 120B model with 12B active parameters, multimodal, intended for code engineering
- Both models include long-context training and native speculative decoding
- The post-training pipeline uses RLVR, RLHF, and for Super, multi-teacher on-policy distillation (MOPD)
- Super achieves 51.78% on SWE-bench Pro and has 4.6× fewer active parameters than Nemotron 3 Ultra
-
The model has 30 billion parameters (specific size); Released under the Linux Foundation OpenMDW 1.1 license with weights and training recipes; Accompanying NeMo Switchyard tool for intelligent routing in agentic applications; Specific applications: code review, tool use, security monitoring, billing questions; Responds to the strengthening of Chinese open-source models (Alibaba Qwen, Moonshot)
- The model has 30 billion parameters (specific size)
- Released under the Linux Foundation OpenMDW 1.1 license with weights and training recipes
- Accompanying NeMo Switchyard tool for intelligent routing in agentic applications
- Specific applications: code review, tool use, security monitoring, billing questions
- Responds to the strengthening of Chinese open-source models (Alibaba Qwen, Moonshot)
Two audiences, two different impacts
What this means
For individuals
Developers and technically oriented users can try an open model with published weights and recipes locally, instead of being limited to paid APIs.
For a business
Companies building agentic applications gain open models with full weights, training data, and recipes, along with documented performance figures, which enables local deployment without reliance on paid APIs and at lower cost under high-volume operation.
DevelopmentCheck the original
Event sources
confirmed by 2 independent sources · 3 publishers, 2 independent. We count feeds from the same owner only once.