NVIDIA has listed the BF16 and NVFP4 variants of its Nemotron 3.5 Lightning 30B-A3B model on Hugging Face, with accompanying evaluation and reproducibility material in the NVIDIA NeMo Gym repository.
NVIDIA has published the NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 model through its official Hugging Face organization. The listing makes a BF16 version of the Nemotron 3.5 Lightning model available to developers evaluating NVIDIA’s language-model family.
The model name identifies it as a 30B-A3B release: a model with 30 billion total parameters and 3 billion active parameters. It is part of NVIDIA’s Nemotron line, which the company presents for generative-AI applications and enterprise language-model workloads.
NVIDIA’s Hugging Face page provides the release-specific material developers need to review before use, including model documentation and applicable usage information. Model cards remain the primary reference for deployment requirements, supported environments, licensing terms, and known limitations.
NVIDIA has separately listed NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 on Hugging Face. According to NVIDIA’s companion model documentation, that release covers the same 30B/3B-active Lightning model for use cases including agents, chatbots, retrieval-augmented generation, and efficiency-oriented deployment.
The BF16 and NVFP4 listings represent different numerical formats within the same named model line. BF16 is commonly used in modern AI training and inference systems, while NVIDIA positions the NVFP4 release around lower-precision, efficiency-focused operation. The practical trade-offs between the variants will depend on a deployment’s hardware, inference software, quantization support, accuracy requirements, and workload shape.
For teams selecting between the releases, a precision label alone is not a sufficient performance indicator. Testing should cover the target environment and the tasks that matter to the organization, such as response quality, tool use, retrieval-grounded answers, latency, throughput, and context-window behavior.
NVIDIA NeMo’s Gym repository also contains reproducibility documentation for the Nemotron 3.5 Lightning release. The document includes material related to the evaluation suite and benchmark recipes used for the model family.
That documentation gives developers a reference point for understanding how NVIDIA structured evaluation of the release. It can also help teams build comparable tests, although results may differ across hardware, inference frameworks, prompts, decoding settings, datasets, and evaluation implementations.
Independent validation remains important for organizations considering the model in production. In addition to general benchmark results, developers may need to measure domain-specific performance, factual reliability, safety behavior, cost, and operational characteristics under expected traffic.
The new Hugging Face listings give developers BF16 and NVFP4 options for evaluating the Nemotron 3.5 Lightning 30B-A3B model. Alongside NVIDIA NeMo’s published reproducibility recipes, the releases provide both model-format choices and supporting evaluation references for teams assessing the model on their own infrastructure.
NVIDIA Adds a BF16 Nemotron 3.5 Lightning Model NVIDIA has published the NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 model through its official Hugging Face organization.
The listing makes a BF16 version of the Nemotron 3.5 Lightning model available to developers evaluating NVIDIA’s language model family.
The model name identifies it as a 30B A3B release: a model with 30 billion total parameters and 3 billion active parameters.
Continue reading
