Nemotron 3 Super Throughput Notes

This article discusses NVIDIA's Nemotron 3 Super 120B-A12B open-weight model, emphasizing its design for balancing accuracy and throughput. It details key efficiency choices like Mamba-2 layers for long-context efficiency, Latent MoE layers for sparse scaling, shared-weight multi-token prediction for speculative decoding, and mixed GQA layers. The model is benchmarked against GPT-OSS 120B and Qwen3.5, showing competitive accuracy but stronger throughput. The article highlights its relevance for local agentic applications where latency and cost are critical, making it a notable development in AI/tech infrastructure.