NVIDIA launches Nemotron 3.5 Lightning: hybrid Mamba-MoE model with 3 billion active parameters

NVIDIA released Nemotron 3.5 Lightning yesterday, a language model with a total of 30 billion parameters that activates only 3 billion per forward pass. The architecture alternates Mamba-2 and Mixture-of-Experts (MoE) layers, supplemented by selected attention layers. Weights, training data and training recipes are provided under the OpenMDW-1.1 license and are available for immediate download on Hugging Face. NVIDIA describes the model as ready for commercial use.

The hybrid design pairs Mamba-2, which offers linear complexity in context length, with MoE, which keeps inference cost constant through sparse expert activation. This combination aims to provide long context windows without the quadratic cost of attention. NVIDIA also adds Multi-Token Prediction (MTP), which predicts several tokens ahead at each step to accelerate speculative text generation. The most recent training data include pre-training up to September 2025 and post-training up to May 2026.

Performance numbers are reported on the NeMo Gym / NeMo Evaluator SDK harness, not through self-reported benchmarks. NVIDIA explicitly notes that the figures may differ from those reported by hardware manufacturers and publishes the full evaluation recipes for reproducibility. Results for SWE-Bench Verified and Terminal-Bench 2.1 are included, and all numbers refer to the official NVFP4 checkpoint.

The quick-start guide targets DGX Spark (GPU GB10) with the nightly vLLM build. The recipe uses a Marlin backend for MoE, FP8 KV cache, prefix caching, speculative decoding with three tokens, a FlashInfer backend for Mamba, and dedicated parsers for inference (nemotron_v3) and tool calls (qwen3_coder). NVIDIA also provides a dedicated checkpoint for DSpark speculative decoding. In short, the model is engineered for efficient execution on current NVIDIA hardware.

OpenMDW-1.1 permits commercial use, a key point for teams building products. The model is intended for English and code, with declared support for Spanish, French, German, Italian and Japanese. NVIDIA positions it for AI agents, chatbots, retrieval-augmented generation and general AI applications, leaving it to the community to compare real-world performance against comparable open models such as Qwen3 and Llama 3.3 outside NVIDIA’s evaluation harness.