nemotron-3.5-lightning-30b-a3b

This hybrid model blends Mamba-2, Mixture-of-Experts, and Attention layers, drawing on NVIDIA's broader Nemotron family of open models while applying NVFP4 quantization to keep inference efficient. With 30B total parameters but only 3B active per token, the architecture is tuned for fast generation through Multi-Token Prediction, pre-trained on over 20T tokens. The 1M-token context window makes it suitable for long-running autonomous agents and sub-agent workhorse roles that need sustained reasoning over extended inputs. Supported languages span English and coding languages, along with Spanish, French, German, Italian, and Japanese, broadening its utility across multilingual developer and enterprise scenarios.

Designed with reasoning and tool calling in mind, the model fits well into agentic pipelines where planning, function use, and multi-step decision-making matter more than conversational polish. NVIDIA highlights reasoning benchmark evaluations on the model card, signaling its intended evaluation against math, coding, and logic suites typical of the Nemotron lineage. The OpenMDW License Agreement v1.1 under which it is released reflects NVIDIA's continued openness for the family, even though the hosted routing through Requesty keeps the weights off the user's machine. For teams building retrieval-heavy or code-execution agents that benefit from a very wide context window, it offers a balanced combination of parameter scale and active-parameter efficiency.