NVIDIA Releases Compressed MoE Model Nemotron-Labs-3-Puzzle-75B-A9B, Boosting Throughput 2.03x

NVIDIA released Nemotron-Labs-3-Puzzle-75B-A9B, a compressed hybrid MoE model from Nemotron-3-Super. Using iterative Puzzle method alternating hardware-aware structure compression and short knowledge distillation recovery, total parameters dropped from 120.7B (active 12.8B) to 75.3B (active 9.3B). On a single 8x B200 node, total throughput reached 2.03x the original Super, while maintaining 100 tok/s per user. On a single H100 GPU, concurrent 1M token requests increased from 1 to 8. For teams deploying large models on NVIDIA GPUs, this means supporting more users or higher concurrency on the same hardware, or halving the number of GPUs needed for equivalent performance.