{"paper":{"title":"SlimQwen: Exploring the Pruning and Distillation in Large MoE Model Pre-training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Pruning a pretrained large MoE consistently outperforms training the smaller target architecture from scratch under the same training budget.","cross_cats":["cs.AI","cs.CL"],"primary_cat":"cs.LG","authors_text":"Bo Zheng, Dayiheng Liu, Liangyu Wang, Rui Men, Shengkun Tang, Siqi Zhang, Xiulong Yuan, Zekun Wang, Zhiqiang Shen, Zihan Qiu","submitted_at":"2026-05-09T06:50:35Z","abstract_excerpt":"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, acr"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across depth, width, and expert compression, pruning a pretrained MoE consistently outperforms training the target architecture from scratch under the same training budget; progressive pruning schedules outperform one-shot compression.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the observed convergence of different one-shot expert compression methods after large-scale continual pretraining will generalize beyond the specific Qwen3-Next-80A3B model and the chosen downstream benchmarks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Pruning pretrained MoE models outperforms training from scratch, different compression methods converge after continued pretraining, and combining KD with language modeling loss plus progressive schedules yields a competitive 23A2B model from Qwen3-Next-80A3B.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Pruning a pretrained large MoE consistently outperforms training the smaller target architecture from scratch under the same training budget.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e7244562dbb9ff2219d209b6c8a080dc63e1c8c334cb18a77d0a968b05b84678"},"source":{"id":"2605.08738","kind":"arxiv","version":2},"verdict":{"id":"e64e0df5-d330-4d2c-8687-08baf1faf45e","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T03:24:19.975833Z","strongest_claim":"Across depth, width, and expert compression, pruning a pretrained MoE consistently outperforms training the target architecture from scratch under the same training budget; progressive pruning schedules outperform one-shot compression.","one_line_summary":"Pruning pretrained MoE models outperforms training from scratch, different compression methods converge after continued pretraining, and combining KD with language modeling loss plus progressive schedules yields a competitive 23A2B model from Qwen3-Next-80A3B.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the observed convergence of different one-shot expert compression methods after large-scale continual pretraining will generalize beyond the specific Qwen3-Next-80A3B model and the chosen downstream benchmarks.","pith_extraction_headline":"Pruning a pretrained large MoE consistently outperforms training the smaller target architecture from scratch under the same training budget."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.08738/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-19T22:35:15.381812Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T14:31:17.561569Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T10:50:48.433480Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"1b83155b8377816cc4861576b945f25ac7ff4ff05efd011084251d8a08e62411"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}