Nemotron 3.5 Lightning 30b A3b

Nemotron 3.5 Lightning 30B A3B is the August 11, 2026 release of NVIDIA's Lightning line, a 30 billion parameter Mixture-of-Experts that activates about 3 billion parameters per token. The backbone interleaves Mamba-2 state-space layers with MoE layers and a small number of attention layers, and NVIDIA points to the NVFP4 release for latency- and throughput-optimized inference. Pretraining ran over more than 20 trillion tokens, and the model card lists a context length of up to 1 million tokens, with 256K used on a single H100. Reasoning can be switched on or off through the chat template, and tool calling is supported out of the box in vLLM and SGLang. NVIDIA reports 81.94 on MMLU Pro, 75.44 on GPQA Diamond, 51.56 on SWE-bench Verified and 24.58 on Terminal-Bench 2.1 for the BF16 reference weights. Beyond English and code, the card lists Spanish, French, German, Italian and Japanese as supported languages, with 43 programming languages in the training data. The weights ship under the OpenMDW-1.1 license in BF16 and NVFP4, with companion speculative-decoding drafters, and the BF16 checkpoint is positioned as the starting point for fine-tuning, distillation and custom quantization. NVIDIA aims it at long-running autonomous agents and sub-agent workhorse deployments.