Nemotron-3-Super-120B-A12B is a large open-weight language model designed for agentic, reasoning, and conversational tasks, including coding, planning, tool calling, and long-context analysis. The model uses a hybrid architecture that interleaves Mamba-2 layers with Mixture-of-Experts layers, and it adds Multi-Token Prediction (MTP) to accelerate generation. A distinctive design choice is the use of LatentMoE, where tokens are projected into a compressed latent space before expert routing, allowing roughly four times more experts to fit within the same parameter budget compared with conventional routing approaches. The model carries 120B total parameters while activating only 12B per forward pass, which is the central efficiency story behind the variant's name and its positioning for high-throughput serving. Multilingual coverage spans English, French, German, Italian, Japanese, Spanish, and Chinese.

Independent benchmarks published shortly after release evaluate the model for API latency and cost across inference providers, indicating that the active-parameter design translates into measurable inference-economics gains relative to dense 120B-class models. The architecture's hybrid Mamba-2 attention mix is well suited to long-context workloads, and the expanded context window supports sustained reasoning and planning tasks that chain multiple tool calls. These characteristics make the model a practical fit for production agents, research workflows that need long document analysis, and developer tooling that benefits from open-weight deployment and fine-tuning freedom. Compared with peers of similar total size, the combination of sparse activation, MTP, and LatentMoE offers a forward-looking path to scaling expert counts without proportional increases in compute per token.