NVIDIA AI has released Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of the Nemotron-3-Super hybrid MoE model. The parent model has 120.7B total and 12.8B active parameters, while the compressed model has 75.3B total and 9.3B active parameters. The deployment target was fixed before the architecture search began, with target one being 2x server throughput at 100 tokens per second per user and target two being 8 concurrent 1M-token requests on a single H100. Three checkpoints are available on Hugging Face: BF16, FP8, and NVFP4. The 88-block hybrid layout is preserved, with 40 Mamba, 40 MoE, and 8 attention blocks. On a single 8xB200 node, total throughput rises 1.60x to 2.14x over Super at matched NVFP4 and matched user throughput. Single-H100 1M-token concurrency goes 1 to 8, driven by a 70 GB to 44.5 GB weight drop. Iterative Puzzle beats single-step Puzzle by 0.57 average points at the same compression target. Benchmark results show Arena-Hard-V2 (-4.2) and SWE-Bench (-2.6) as the real costs, while RULER and AA-LCR barely move.