Nvidia has expanded the availability of Nemotron 3.5 Lightning to the Hugging Face platform, allowing users to run the model through the popular AI platform. The 31.6 billion parameter model with 3.6 billion active parameters per token is specifically designed for agentic tasks such as reading files, calling tools, sorting results, and retrying failed operations. The model excels on workloads where throughput is the critical metric rather than large-scale computation, making it ideal for local systems rather than datacenter deployment. Users can now access the model through Hugging Face's Transformers library and vLLM framework, providing seamless integration with existing AI development workflows.
The model employs a hybrid Latent Mixture-of-Experts (LatentMoE) architecture utilizing interleaved Mamba-2 and MoE layers, along with select Attention layers, according to NVIDIA. This innovative approach allows the model to handle the capacity of a larger model while significantly reducing computation requirements. The model has 30 billion total parameters with 3 billion active parameters and supports up to 1 million token context length. The company reports that Nemotron 3.5 Lightning reached 24 on the Artificial Analysis Intelligence Index, representing a +9 point improvement over Nemotron 3 Nano (15). Nvidia claims the model can deliver up to four times the output speed of similarly sized models through speculative decoding and the inclusion of two draft models, DSpark and DFlash. The model offers 4x higher throughput and 30% faster task completion time compared to other leading open models of similar size.
Nvidia has expanded the Nemotron 3.5 Lightning ecosystem with local deployment capabilities across NVIDIA RTX PCs, NVIDIA DGX Spark, OEM GB10 systems, and NVIDIA Jetson, as reported by NVIDIA. The model runs on NVIDIA Blackwell systems from Acer, ASUS, Dell Technologies, Exxact, GIGABYTE, HP, Lenovo, MSI and Supermicro, providing users with choice of form factors. Nvidia collaborated with vLLM, Ollama, llama.cpp and LM Studio to provide the best local deployment experience, offering developers choice of NVFP4 and GGUF format of models. Unsloth provides day-one support with optimized and quantized models for efficient local deployment via Unsloth Studio. For Apple silicon users, Ollama offers the model with state-of-the-art performance through nemotron-3.5-lightning:30b-mlx configuration. The model is fully open—weights, data, and recipes—so you can adapt it to your workflows and deploy it anywhere.
The Nemotron 3.5 Lightning ecosystem is supported by NVIDIA NeMo Switchyard, a library that intelligently routes each task to the best model for the job. Switchyard can expose Nemotron 3.5 Lightning as a routing target alongside your open and closed models, so every request lands on the most capable and efficient model that can handle it. Plans route up to the frontier, execution routes down to Lightning, ensuring that your tokens are spent efficiently and effectively. The model is designed to work alongside larger reasoning models like Nemotron 3 Ultra for planning and orchestration, while Lightning handles high-volume execution once agents have determined what tasks need completion. The platform allows agents to send individual steps to larger models without changing anything else, maintaining seamless integration across different model sizes and deployment environments.
The launch aligns with Nvidia's broader coalition supporting open-weight AI models that has urged policymakers to avoid broad restrictions on models whose weights can be downloaded and modified, according to Business Standard. Nvidia CEO Jensen Huang, along with Meta and other technology companies, formed this coalition to support the growing industry push for open AI models. The company is releasing the model with weights, training data and recipes under the OpenMDW-1.1 license, allowing developers to fine-tune it using Nvidia's NeMo tools and use reinforcement learning and environment-based evaluations. The model's availability on Hugging Face platform further strengthens its position in the open AI ecosystem. This release includes Nemotron-RL Agentic Terminal Pivot, an open agentic reinforcement learning dataset used to train some of the coding agent capabilities.
