Nvidia has officially launched Nemotron 3 Super, a 120-billion-parameter open-source hybrid model that activates only 12.7 billion parameters per forward pass through a Mixture-of-Experts (MoE) architecture. The model is designed to reduce the cost of running AI agents at scale, targeting the ballooning token usage and extended reasoning chains that plague multi-step agent pipelines.
Architectural Innovations: Mamba-Transformer Hybrid and LatentMoE
Nemotron 3 Super employs an 88-layer hybrid backbone combining Mamba-2 blocks with Transformer attention layers. The Mamba-2 modules handle long sequences with linear-time efficiency, while the Transformer layers retain precise recall capabilities. This blend enables native support for context windows of up to one million tokens without the memory overhead typical of pure-attention models.
The model's MoE system, called LatentMoE, compresses token embeddings into a low-rank space before routing them to 512 experts per layer, activating 22 at a time. Nvidia claims this allows roughly four times more experts at the same inference cost compared to standard MoE, enabling finer task specialization—for example, separating Python logic from SQL handling at the expert level.
Multi-Token Prediction layers, using two shared-weight heads, accelerate chain-of-thought generation and enable native speculative decoding. On structured tasks, Nvidia reports up to three times faster generation.
Training Data and Benchmark Performance
Nemotron 3 Super was pre-trained on 25 trillion tokens in two phases: the first phase used 20 trillion tokens of broad data, and the second phase used 5 trillion high-quality tokens to fine-tune benchmark performance. A final extension phase on 51 billion tokens extended native context to one million tokens. Post-training included supervised fine-tuning on approximately 7 million samples and reinforcement learning across 21 environments with more than 1.2 million rollouts.
In benchmarks, the model scored 83.73 on MMLU-Pro, 90.21 on AIME25, and 60.47 on SWE-Bench (using OpenHands). On PinchBench, it reached 85.6%, the highest reported score among open models in its class. On long-context evaluation, it scored 91.64 on RULER 1M.
Compared to GPT-OSS-120B, Nemotron 3 Super delivers 2.2x throughput at 8k input and 64k output. Against Qwen3.5-122B-A10B, that figure reaches 7.5x. Nvidia also reports more than 5x throughput and up to 2x accuracy improvements over the prior Nemotron Super generation.
Hardware Optimization and Full-Stack Training
Nvidia trained the model end-to-end in its NVFP4 four-bit floating-point format, optimized for Blackwell GPUs. On B200 hardware, inference runs up to four times faster compared to FP8 on H100 with no reported accuracy loss. Quantized FP8 and NVFP4 checkpoints retain 99.8% or more of full-precision accuracy.
The model also powers the Nvidia AI-Q research agent, which achieved the top position on the Deepresearch Bench leaderboard.
Open Source and Deployment
Nemotron 3 Super is fully open under the Nvidia Nemotron Open Model License. Checkpoints in BF16, FP8, and NVFP4 formats, along with pre-training data, post-training samples, and reinforcement learning environments, are available on Hugging Face. Inference is supported through Nvidia NIM, build.nvidia.com, Perplexity, Openrouter, Together AI, Google Cloud, AWS, Azure, and Coreweave, with on-premises options via Dell Enterprise Hub and HPE.
Developers can access training recipes, fine-tuning guides, and inference cookbooks through the NeMo platform using vLLM, SGLang, and TensorRT-LLM.