Nvidia drops Nemotron 3 Super to supercharge AI agents

Nvidia has open-sourced Nemotron 3 Super, a 120B-parameter hybrid model designed specifically to power complex, multi-agent systems at scale. Despite its massive total capacity, the model activates just 12B parameters at inference time and natively supports a 1-million-token context window.

The model is built to solve two massive bottlenecks holding autonomous agents back: context explosion and the thinking tax. Multi-agent workflows often generate up to 15x more tokens than standard chats because every interaction requires resending conversation history, tool outputs, and reasoning steps. Nemotron 3 Super cuts through that drag, delivering up to 5x higher throughput and up to 2x higher accuracy than its predecessor.

Under the hood

The efficiency comes down to a clever combination of three architectural changes.

First is a hybrid Mamba-Transformer backbone. Most sequence processing is handled by Mamba-2 layers, which offer linear computational complexity and deliver 4x higher memory and compute efficiency. Transformer attention layers are strategically interleaved at key depths to preserve precise associative recall across long sequences.

Second is a new technique called Latent MoE. Instead of routing tokens through full hidden dimensions, the model compresses them into a low-rank latent space first. This lets the system call four specialist experts for the computing cost of one.

Finally, multi-token prediction allows the model to predict multiple future tokens in a single forward pass, boosting raw generation speed by 3x. On Nvidia Blackwell systems running native NVFP4 precision, inference runs up to 4x faster than Hopper FP8 without sacrificing accuracy.

Performance and open release

Nemotron 3 Super currently tops the Artificial Analysis leaderboard for efficiency and openness among models in its weight class. It powered Nvidia's AI-Q research agent to the top of both DeepResearch Bench and DeepResearch Bench II, and scored 85.6% across the full PinchBench suite evaluating OpenClaw agent backbones.

Nvidia released the model with open weights under a permissive license. The release includes the complete training methodology, covering more than 10 trillion tokens of pre- and post-training data recipes and 15 reinforcement learning environments.

Developers can test Nemotron 3 Super today on build.nvidia.com, access it via Hugging Face and OpenRouter, or run it through Perplexity. For production, the model is packaged as an Nvidia NIM microservice for on-premise hardware, Dell and HPE agent stacks, and major cloud providers including Google Cloud Vertex AI and OCI.

Why it matters

If you are building software engineering agents or autonomous security triage pipelines, Nemotron 3 Super deserves immediate attention. Loading an entire repository into a million-token window without blowing up your compute bill finally makes long-horizon agent execution practical.

Strategically, this is classic Nvidia. By releasing an open, state-of-the-art agent model that runs drastically faster on Blackwell hardware with NVFP4, Nvidia is creating a powerful hardware incentive while giving developers a serious open alternative to proprietary frontier APIs.