30B total / 3B active (MoE)
1,048,576 tokens (1M; 256K used for single-H100 BF16 deployment)
OpenMDW License Agreement, version 1.1
Open-weight general-purpose reasoning and agentic model built on NVIDIA's Nemotron-3-Lightning architecture. Nemotron 3.5 Lightning uses a sparse hybrid Mixture-of-Experts design with interleaved Mamba-2, MoE, and selected attention layers, activating 3B of its 30B parameters per token. It was pretrained on more than 20 trillion tokens and includes Multi-Token Prediction heads; subsequent supervised training covered code, mathematics, science, tool calling, structured output, instruction following, long-range retrieval, and multi-document aggregation, followed by asynchronous multi-environment GRPO reinforcement learning across reasoning, multi-step tool use, multi-turn conversation, and coding environments. The model supports up to a 1M-token context window, automatic schema-based tool calling, switchable thinking and non-thinking modes, a runtime reasoning budget, and English, Spanish, French, German, Italian, Japanese, and programming languages. NVIDIA positions the optimized release for long-running autonomous agents, local subagent deployments, chatbots, and RAG systems. Official variants include full-precision BF16 reference and base weights plus an NVFP4 checkpoint that also supports W4A16 serving on Ampere hardware; a GGUF conversion is published through ggml-org. The release supports vLLM, SGLang, TensorRT-LLM, and speculative decoding through native MTP plus external DSpark semi-autoregressive and DFlash block-diffusion draft models. NVIDIA reports BF16 scores of 81.94 on MMLU-Pro, 75.44 on GPQA Diamond, 51.56 on SWE-Bench Verified, 36.97 on BrowseComp, 85.37 on PinchBench, and 832 Elo on GDPval-AA-V2; the NVFP4 variant scores 52.80 on SWE-Bench Verified, 36.81 on BrowseComp, and 865 Elo on GDPval-AA-V2. NVIDIA also published reproducible NeMo Gym and NeMo Evaluator recipes for its evaluation suite.