# NVIDIA 将 Nemotron-3-Super 压缩为 Puzzle-75B-A9B，吞吐量翻倍

- 来源：elvis (@omarsar0)
- 发布时间：2026-07-08 01:20
- AIHOT 分数：52
- AIHOT 链接：https://aihot.virxact.com/items/cmraxp6uq01ynihogtvyg7s4t
- 原文链接：https://x.com/omarsar0/status/2074543978129793462

## AI 摘要

NVIDIA 将混合 MoE 大模型 Nemotron-3-Super 压缩为 Puzzle-75B-A9B。通过联合结构搜索同时优化异构 MoE 剪枝、活跃参数预算和 Mamba 剪枝，配合蒸馏、强化学习、量化及多 token 预测头的迭代流程，在保持质量的同时将交互式服务器吞吐量提升约 2 倍。在单台 8×B200 节点上达约 2 倍父模型吞吐量，单个 H100 上 1M token 并发从 1 请求提升至 8 请求。推理、编码、长上下文和智能体基准准确率不变。论文：arxiv.org/abs/2607.04371。

## 正文

Banger compression paper from NVIDIA.

（bookmark it）

Bigger MoE models keep winning on quality， but serving them at interactive latency is still hard.

NVIDIA compresses the hybrid MoE Nemotron-3-Super into Puzzle-75B-A9B and roughly doubles interactive server throughput while holding quality.

Pay attention to the joint structural search. Heterogeneous MoE pruning， active-parameter budget， and Mamba pruning get optimized together rather than one at a time， wrapped in an iterative pipeline with distillation， RL， quantization， and a Multi-Token Prediction head.

Why does it matter？

On a single 8xB200 node it hits about 2x the parent's server throughput at matched user-throughput， and 1M-token concurrency on a single H100 climbs from 1 request to 8. Accuracy holds across reasoning， coding， long-context， and agentic benchmarks.

Cheaper serving with agentic capability intact changes what you can afford to run with these models.

Paper： https://arxiv.org/abs/2607.04371

Learn to build effective AI agents in our academy： https://academy.dair.ai/