Nvidia has introduced Nemotron 3 Super, a new open hybrid model aimed at large-scale agentic AI workloads, where long reasoning chains and multi-agent pipelines can sharply increase compute costs. While the model contains 120 billion total parameters, Nvidia says it activates only 12.7 billion parameters per forward pass, using a Mixture-of-Experts design to reduce inference overhead without sacrificing capability.
The release positions Nemotron 3 Super as the second model in Nvidia’s Nemotron 3 family, following Nemotron 3 Nano. The company says the model was built to directly address a growing challenge in production AI systems: agents that consume significantly more tokens and compute because they reason over multiple steps, call tools, and coordinate across workflows. In those settings, token usage can multiply dramatically, making throughput and cost efficiency central concerns for developers and infrastructure providers.
Hybrid Mamba-Transformer Design for Long Context
At the architecture level, Nemotron 3 Super uses an 88-layer hybrid Mamba-Transformer backbone. Nvidia says the Mamba-2 blocks help process long sequences with linear-time efficiency, while Transformer attention layers preserve precise recall. The result is native support for context windows of up to 1 million tokens, a capability Nvidia frames as a way to avoid the memory penalties often associated with pure-attention systems at very long context lengths.
The model also incorporates a LatentMoE routing system. In this setup, token embeddings are first compressed into a low-rank latent space before being routed across 512 experts per layer, with only 22 experts activated at a time. According to Nvidia, this approach allows for roughly four times more experts at the same inference cost as standard MoE methods, enabling finer specialization across tasks. The company gives examples such as separating Python logic handling from SQL-oriented processing at the expert level.
Performance Claims Focus on Throughput and Structured Tasks
Nvidia says Nemotron 3 Super includes multi-token prediction layers with two shared-weight heads, a design intended to speed up chain-of-thought generation and enable native speculative decoding. On structured tasks, the company reports generation speedups of up to 3x.
In benchmark disclosures, Nemotron 3 Super scored 83.73 on MMLU-Pro, 90.21 on AIME25, and 60.47 on SWE-Bench using OpenHands. On PinchBench, it reached 85.6%, which Nvidia describes as the highest recorded score among open models in its class. In long-context evaluation, the model posted 91.64 on RULER 1M.
Nvidia also highlighted throughput comparisons under an 8k input / 64k output configuration. Under those conditions, the company says Nemotron 3 Super delivers 2.2x higher throughput than GPT-OSS-120B and as much as 7.5x higher throughput than Qwen3.5-122B-A10B. It further claims more than 5x higher throughput and up to 2x better accuracy than the previous-generation Nemotron Super.
Massive Training Pipeline and Blackwell Optimization
The model was pre-trained on a total of 25 trillion tokens across two stages. The first used 20 trillion tokens of general-purpose data, while the second used 5 trillion tokens of higher-quality data aimed at improving benchmark performance. Nvidia then carried out a final context-extension phase on 51 billion tokens, expanding the model’s native context length to 1 million tokens.
For post-training, Nvidia says it applied supervised fine-tuning on approximately 7 million samples and reinforcement learning across 21 environments with more than 1.2 million rollouts. Those figures reflect a large-scale optimization effort geared toward both reasoning quality and practical agent behavior.
Nemotron 3 Super was trained end-to-end in Nvidia’s NVFP4 four-bit floating-point format, which is optimized for Blackwell GPUs. On B200 hardware, Nvidia claims inference is up to 4x faster than FP8 on H100, with no reported loss in accuracy. The company also says quantized FP8 and NVFP4 checkpoints preserve 99.8% or more of full-precision accuracy.
Nvidia added that the model powers its AI-Q research agent, which reached the top spot on the Deepresearch Bench leaderboard.
Open Release and Broad Deployment Options
A notable part of the announcement is the release model. Nemotron 3 Super is available under the Nvidia Nemotron Open Model License, with checkpoints in BF16, FP8, and NVFP4 formats. Nvidia says pre-training data, post-training samples, and reinforcement learning environments are also available on Hugging Face, giving developers deeper visibility into how the system was built.
Deployment support spans both cloud and enterprise environments. Nvidia says inference is supported through Nvidia NIM, build.nvidia.com, Perplexity, Openrouter, Together AI, Google Cloud, AWS, Azure, and Coreweave. For on-premise users, options include Dell Enterprise Hub and HPE. Developers can also access training recipes, fine-tuning guides, and inference cookbooks through the NeMo platform using tools such as vLLM, SGLang, and TensorRT-LLM.
Overall, the launch of Nemotron 3 Super reflects Nvidia’s broader push to make high-performance, long-context, agent-focused AI models more practical to run at scale. The company’s message is clear: in the next phase of AI competition, raw model size matters less if architecture, routing efficiency, and hardware optimization can deliver significantly better throughput per unit of compute.