NVIDIA's Nemotron 3 Super in 6 Minutes

A New Take on Mixture of Experts

NVIDIA released Nemotron 3 Super, and the architecture is worth paying attention to. It is a 120B parameter mixture-of-experts model, but only about 12B parameters are active per token. That ratio alone makes it interesting for inference costs. What makes it different from standard MoE is the "latent" approach - instead of routing raw tokens to experts, the model compresses tokens into a smaller representation before routing. Experts process these compressed inputs, which means you can run up to four times more experts at the same computational cost as a traditional MoE setup.

The other architectural piece is the hybrid Mamba integration. NVIDIA blends transformer attention layers with Mamba state-space layers, getting transformer-quality reasoning with Mamba's linear scaling on long sequences. The result is a model that handles its full 1M token context window efficiently, especially in multi-user serving scenarios where throughput matters more than single-request latency.

Openness Done Right

One of the more notable aspects of Nemotron 3 Super is how NVIDIA handled the release. You can download the weights, self-host, fine-tune, and commercialize. The training documentation is published. This is the kind of openness that actually matters for developers - not just a model card and an API endpoint, but the full package that lets you build on top of it.

NVIDIA positions this as a balance between openness and capability. Many open models sacrifice intelligence for permissive licensing, or gate the best checkpoints behind restrictive terms. Nemotron 3 Super ships competitive benchmarks alongside genuinely permissive access. For teams evaluating sub-250B models for production use, that combination narrows the field significantly.

Where to Run It

The model is available today through several channels. Perplexity has it integrated. Hugging Face hosts the weights for self-hosting. Major cloud providers offer managed inference. NVIDIA's own developer tools and build platform provide direct access for testing before you commit to infrastructure.

Benchmark results show improved throughput and coding performance versus prior Nemotron releases and other models in the sub-250B class. The latent MoE architecture pays off most visibly in multi-user scenarios - the compressed expert routing means you serve more concurrent requests before hitting memory or compute ceilings. For teams running inference at scale, the 12B active parameter footprint per token translates directly to lower cost per query while maintaining the quality of a much larger model.

Check out the full breakdown in the video above, or grab the weights from Hugging Face and try it yourself.