NVIDIA released Nemotron 3.5 Lightning on August 11, 2026, and it arrives with a clear purpose: to serve as the high-volume execution layer for long-running AI agents without demanding a multi-GPU cluster to run. It is a 30-billion-parameter model that activates only 3 billion parameters per token, built for developers who need speed, long context, and tool-calling reliability in a package that fits on a single accelerator.
A Hybrid Architecture Built for Efficiency
The architecture is a hybrid Mixture-of-Experts design that interleaves Mamba-2 state-space layers with MoE layers and select Attention layers. This combination is what allows the model to keep inference costs low while still handling tasks that require reasoning and long-range retrieval. The MoE configuration uses 128 routed experts with 6 activated per token, and the model spans 52 layers in total. NVIDIA also built in Multi-Token Prediction layers, which provide richer training signals and native speculative decoding for faster generation.
Context length reaches up to 1 million tokens, with a native maximum of 262,144 positions. That matters for agentic workflows where the model must keep track of tool outputs, conversation history, and retrieved documents across extended sessions. The model supports English and coding languages along with Spanish, French, German, Italian, and Japanese.
Benchmarks: Strong for Its Size Class
On the BF16 checkpoint, Nemotron 3.5 Lightning scores 81.94 on MMLU Pro, 75.44 on GPQA Diamond without tools, and 51.56 on SWE-bench Verified. The NVFP4 variant comes in at 81.62 on MMLU Pro, 75.57 on GPQA Diamond, and 52.80 on SWE-bench Verified. On long-context retrieval, the model reaches 52.00 on AA-LCR, and on instruction following it scores 71.88 on IFBench. Compared directly with other models in the 30B range, it outperforms the Nemotron 3 Nano 30B-A3B across MMLU, MMLU-Pro, AGIEval-EN, and GSM8K.
Deployment: One GPU Is Enough
The BF16 weights take up roughly 60GB of VRAM, which means a single 80GB H100 or A100 can serve the model with headroom for the KV cache. The NVFP4 checkpoint drops the memory footprint to the mid-teens of gigabytes, opening the door to a GeForce RTX 5090, a DGX Spark (GB10), or older Ampere-class cards running W4A16 kernels. For serving, vLLM 0.27.1 or newer is recommended, since the hybrid Mamba-2 architecture requires the Mamba serving path that recent vLLM builds add. The model is also available through NVIDIA NIM, Ollama, OpenRouter, and Amazon SageMaker JumpStart.
Open License, Commercial Use
Nemotron 3.5 Lightning ships under the OpenMDW License Agreement, version 1.1, which permits unrestricted commercial use, modification, and distribution. The base checkpoint is also available for developers who want to apply their own supervised fine-tuning, reinforcement learning, or distillation without starting from a post-trained model. The pre-training data has a cutoff of September 2025, and the post-training data extends through May 2026.
For teams building agent systems that need to run continuously, call tools reliably, and stay affordable at scale, Nemotron 3.5 Lightning is a deliberately practical release. It does not try to be the largest model in the room. It tries to be the one that fits on the hardware you already have.
Email your news TIPS to Editor@Kahawatungu.com — this is our only official communication channel

