SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training

• URL: http://arxiv.org/abs/2605.08738v2
• Date: Mon, 18 May 2026 06:29:11 GMT
• Title: SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training
• Abstract summary: We systematically study MoE compression in large-scale pretraining.<n>We find that pruning a pretrained MoE consistently outperforms training the target architecture from scratch.<n>We propose multi-token prediction (MTP) distillation, which yields consistent gains.
• Score: 57.41616809842774
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Structured pruning and knowledge distillation (KD) are typical techniques for compressing large language models, but it remains unclear how they should be applied at pretraining scale, especially to recent mixture-of-experts (MoE) models. In this work, we systematically study MoE compression in large-scale pretraining, focusing on three key questions: whether pruning provides a better initialization than training from scratch, how expert compression choices affect the final model after continued training, and which training strategy is most effective. We have the following findings: First, across depth, width, and expert compression, pruning a pretrained MoE consistently outperforms training the target architecture from scratch under the same training budget. Second, different one-shot expert compression methods converge to similar final performance after large-scale continual pretraining. Motivated by this, we introduce a simple partial-preservation expert merging strategy that improves downstream performance across most benchmarks. Third, combining KD with the language modeling loss outperforms KD alone, particularly on knowledge-intensive tasks. We further propose multi-token prediction (MTP) distillation, which yields consistent gains. Finally, given the same training tokens, progressive pruning schedules outperform one-shot compression, suggesting that gradual architecture transitions lead to better optimization trajectories. Putting it all together, we compress Qwen3-Next-80A3B to a 23A2B model that retains competitive performance. These results offer practical guidance for efficient MoE compression at scale.

Related papers

• A Systematic Study of Compression Ordering for Large Language Models [0.5926203312586109]
  This study systematically examines how knowledge distillation, structured pruning, and low-bit quantization perform when applied to the Qwen2.5 3B model.<n>Experiments show that quantization provides the greatest standalone compression, while pruning introduces moderate quality degradation.
  arXiv  Detail & Related papers  (2025-11-23T12:46:56Z)
• Compressing Many-Shots in In-Context Learning [61.231471139896506]
  We study an approach to improve the memory and computational efficiency of ICL inference by compressing the many-shot prompts.<n>We first show that existing prompt compression methods are ineffective for many-shot compression.<n>We propose MemCom, a layer-wise compression method.
  arXiv  Detail & Related papers  (2025-10-17T16:57:42Z)
• Overtrained Language Models Are Harder to Fine-Tune [64.44743256512237]
  Large language models are pre-trained on ever-growing token budgets.<n>We show that extended pre-training can make models harder to fine-tune, leading to degraded final performance.
  arXiv  Detail & Related papers  (2025-03-24T23:11:56Z)
• MultiPruner: Balanced Structure Removal in Foundation Models [1.8434042562191815]
  Recently, state-of-the-art approaches for pruning large pre-trained models (LPMs) have demonstrated that the training-free removal of non-critical residual blocks in Transformers is viable for reducing model size.<n>We extend BlockPruner and propose MultiPruner, a pruning approach that surpasses recent training-free pruning methods by adopting a multidimensional, iterative, fine-grained pruning strategy.
  arXiv  Detail & Related papers  (2025-01-17T04:24:31Z)
• Activations and Gradients Compression for Model-Parallel Training [85.99744701008802]
  We study how simultaneous compression of activations and gradients in model-parallel distributed training setup affects convergence. We find that gradients require milder compression rates than activations. Experiments also show that models trained with TopK perform well only when compression is also applied during inference.
  arXiv  Detail & Related papers  (2024-01-15T15:54:54Z)
• Retraining-free Model Quantization via One-Shot Weight-Coupling Learning [41.299675080384]
  Mixed-precision quantization (MPQ) is advocated to compress the model effectively by allocating heterogeneous bit-width for layers. MPQ is typically organized into a searching-retraining two-stage process. In this paper, we devise a one-shot training-searching paradigm for mixed-precision model compression.
  arXiv  Detail & Related papers  (2024-01-03T05:26:57Z)
• Just CHOP: Embarrassingly Simple LLM Compression [27.64461490974072]
  Large language models (LLMs) enable unparalleled few- and zero-shot reasoning capabilities but at a high computational footprint. We show that simple layer pruning coupled with an extended language model pretraining produces state-of-the-art results against structured and even semi-structured compression of models at a 7B scale. We also show how distillation, which has been super effective in task-agnostic compression of smaller BERT-style models, becomes inefficient against our simple pruning technique.
  arXiv  Detail & Related papers  (2023-05-24T08:18:35Z)
• CrAFT: Compression-Aware Fine-Tuning for Efficient Visual Task Adaptation [3.043665249713003]
  Post-training compression techniques such as pruning and quantization can help lower deployment costs. We propose CrAFT, a simple fine-tuning framework that enables effective post-training network compression. The CrAFT approach adds negligible training overhead as fine-tuning is done in under a couple of minutes or hours with a single GPU.
  arXiv  Detail & Related papers  (2023-05-08T07:51:40Z)
• Does compressing activations help model parallel training? [64.59298055364336]
  We present the first empirical study on the effectiveness of compression methods for model parallelism. We implement and evaluate three common classes of compression algorithms. We evaluate these methods across more than 160 settings and 8 popular datasets.
  arXiv  Detail & Related papers  (2023-01-06T18:58:09Z)
• Gradient-based Intra-attention Pruning on Pre-trained Language Models [21.444503777215637]
  We propose a structured pruning method GRAIN (Gradient-based Intra-attention pruning) GRAIN inspects and prunes intra-attention structures, which greatly expands the structure search space and enables more flexible models. Experiments on GLUE, SQuAD, and CoNLL 2003 show that GRAIN notably outperforms other methods, especially in the high sparsity regime.
  arXiv  Detail & Related papers  (2022-12-15T06:52:31Z)
• Adversarial Robustness: From Self-Supervised Pre-Training to Fine-Tuning [134.15174177472807]
  We introduce adversarial training into self-supervision, to provide general-purpose robust pre-trained models for the first time. We conduct extensive experiments to demonstrate that the proposed framework achieves large performance margins.
  arXiv  Detail & Related papers  (2020-03-28T18:28:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.