NVIDIA’s Nemotron 3 Super just landed on Perplexity — and it’s a big deal for anyone building AI agents or using the platform for complex tasks.

Announced on March 13, Perplexity confirmed that NVIDIA’s freshly released Nemotron 3 Super model is now available across three of its surfaces: the main Perplexity app, the Agent API, and Perplexity Computer. The rollout comes just days after NVIDIA officially introduced the model on March 11, making its integration into Perplexity one of the fastest post-launch deployments we’ve seen from the platform.

So what exactly is Nemotron 3 Super? It’s a 120-billion-parameter open-weight model — but with a clever twist. Only 12 billion of those parameters are actually active at any given time during inference, thanks to its hybrid Mixture-of-Experts (MoE) architecture. Think of it like having a team of 120 specialists, but only waking up the 12 most relevant ones for each task. This makes it dramatically more efficient to run than a dense model of comparable total size.

Under the hood, Nemotron 3 Super blends Mamba layers (great for speed and long sequences) with Transformer layers (great for precise reasoning), a combination NVIDIA calls a hybrid Mamba-Transformer backbone. It also supports a massive 1-million-token context window, meaning it can hold an entire codebase or lengthy research document in memory without losing track. NVIDIA claims it delivers up to 5× higher throughput compared to earlier Nemotron models, and it runs particularly well on NVIDIA’s Blackwell GPUs.

The model is specifically designed for agentic AI workloads — not casual chatting. This means it excels at multi-step planning, tool calling, code writing, and complex reasoning tasks like software development and cybersecurity triaging. It was trained on over 10 trillion tokens of pre-training and post-training data, and NVIDIA is releasing not just the weights but also the training data methodology and recipes, making it a genuinely open model for developers.

That’s why the Perplexity Computer integration is especially exciting. Perplexity Computer is the company’s autonomous agent system that uses 19+ AI models to execute complex, multi-step workflows — from drafting emails and building apps to running parallel research across seven search types simultaneously. Adding Nemotron 3 Super into that mix gives Computer another powerful engine optimized exactly for the kind of agentic, multi-step tasks it was designed to handle. It joins a lineup that already includes Claude Opus 4.6 for core reasoning, Gemini for deep research, and Grok for speed.

For developers, Nemotron 3 Super is also accessible through Perplexity’s Agent API, part of the platform’s recently expanded four-API suite that includes Search, Agent, Embeddings, and Sandbox. As one observer on X noted, integrating Nemotron 3 Super into agent workflows is “going to be a massive unlock for complex reasoning tasks” — and given the model’s architecture, that’s not just hype.

The model is also available for deployment through cloud partners like CoreWeave, Together AI, and Nebius, and as an NVIDIA NIM microservice for on-premise deployments. It’s open-weight, so developers can download the checkpoints, datasets, and training recipes directly.​

In short, Nemotron 3 Super isn’t trying to be your everyday chatbot — it’s built for the messy, multi-layered tasks that autonomous AI agents need to tackle. And with Perplexity now offering it front and center, that capability just became a lot more accessible.