{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:3ZDXG3WWLTJRVGGWPMP2XHXORU","short_pith_number":"pith:3ZDXG3WW","schema_version":"1.0","canonical_sha256":"de47736ed65cd31a98d67b1fab9eee8d19ba2b8bf3ed48f33e893bb36f6e1ce9","source":{"kind":"arxiv","id":"2506.07900","version":2},"attestation_state":"computed","paper":{"title":"MiniCPM4: Ultra-Efficient LLMs on End Devices","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Ao Sun, Baoxi Ji, Bingxiang He, Biyuan Lin, Chao Jia, Chuyue Zhou, Cunliang Kong, Dahai Li, Dan Liu, Fangzheng Wang, Feng Wang, Ganqu Cui, Ge Zhou, Guoyang Zeng, Haotian Chen, Hengyu Zhao, Hongya Lyu, Jiarui Yuan, Jiayuan Su, Jie Cai, Jie Xie, Jie Zhou, Jinqian Zhang, Junshao Guo, Kaihuo Zhang, Lei Zhang, Linyue Zhang, Litu Ou, Lushi Pu, Maosong Sun, MiniCPM Team: Chaojun Xiao, Ning Ding, Peijun Tang, Peiyan Luo, Qiuzuo Li, Quanyu Lu, Qundong Shi, Shengda Fan, Shuo Wang, Siyuan Li, Weilin Zhao, Weilun Zhao, Wei Zhou, Wenhao Li, Wentong Chen, Wenyu Guan, Xianghui Sun, Xiang Long, Xiaoyue Xu, Xin Cong, Xin Li, Xueren Zhang, Xu Han, Yanghao Li, Yanghao Zhou, Yankai Lin, Yaxi Lu, Yesai Wu, Yewei Fang, Yinxu Pan, Yishan Li, Yitong Guan, Yuanqian Zhao, Yudi Zhang, Yudong Wang, Yufeng Han, Yukun Yan, Yuxiang Huang, Yuxuan Li, Yuzhuo Bai, Zekai Qu, Zheng Wang, Zhen Li, Zhenyu Xiao, Zhiyuan Liu, Zhi Zheng, Zhou Su, Zihan Zhou, Zihao Xie, Zijun Song, Zixuan Fu, Zixuan Zhou","submitted_at":"2025-06-09T16:16:50Z","abstract_excerpt":"This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and "},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2506.07900","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2025-06-09T16:16:50Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"b2aa2f73e95bee2720e28242b26283c70636ce21cbdca3d2dd5d7a691f419a14","abstract_canon_sha256":"ae1358eb261905249e74425e0832a79663ffcb076ec10f4a29051bb21b4dbd30"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T12:04:36.204680Z","signature_b64":"PhqT/09b0Q7jkYmxQ4D6kRWw9T88TGx1vcq4wku/LMdvaTnfPHfifiqYaVh4SqUJ2n5oQiX7SAKNHwamJVPJCg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"de47736ed65cd31a98d67b1fab9eee8d19ba2b8bf3ed48f33e893bb36f6e1ce9","last_reissued_at":"2026-07-05T12:04:36.204160Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T12:04:36.204160Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"MiniCPM4: Ultra-Efficient LLMs on End Devices","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Ao Sun, Baoxi Ji, Bingxiang He, Biyuan Lin, Chao Jia, Chuyue Zhou, Cunliang Kong, Dahai Li, Dan Liu, Fangzheng Wang, Feng Wang, Ganqu Cui, Ge Zhou, Guoyang Zeng, Haotian Chen, Hengyu Zhao, Hongya Lyu, Jiarui Yuan, Jiayuan Su, Jie Cai, Jie Xie, Jie Zhou, Jinqian Zhang, Junshao Guo, Kaihuo Zhang, Lei Zhang, Linyue Zhang, Litu Ou, Lushi Pu, Maosong Sun, MiniCPM Team: Chaojun Xiao, Ning Ding, Peijun Tang, Peiyan Luo, Qiuzuo Li, Quanyu Lu, Qundong Shi, Shengda Fan, Shuo Wang, Siyuan Li, Weilin Zhao, Weilun Zhao, Wei Zhou, Wenhao Li, Wentong Chen, Wenyu Guan, Xianghui Sun, Xiang Long, Xiaoyue Xu, Xin Cong, Xin Li, Xueren Zhang, Xu Han, Yanghao Li, Yanghao Zhou, Yankai Lin, Yaxi Lu, Yesai Wu, Yewei Fang, Yinxu Pan, Yishan Li, Yitong Guan, Yuanqian Zhao, Yudi Zhang, Yudong Wang, Yufeng Han, Yukun Yan, Yuxiang Huang, Yuxuan Li, Yuzhuo Bai, Zekai Qu, Zheng Wang, Zhen Li, Zhenyu Xiao, Zhiyuan Liu, Zhi Zheng, Zhou Su, Zihan Zhou, Zihao Xie, Zijun Song, Zixuan Fu, Zixuan Zhou","submitted_at":"2025-06-09T16:16:50Z","abstract_excerpt":"This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2506.07900","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2506.07900/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2506.07900","created_at":"2026-07-05T12:04:36.204221+00:00"},{"alias_kind":"arxiv_version","alias_value":"2506.07900v2","created_at":"2026-07-05T12:04:36.204221+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2506.07900","created_at":"2026-07-05T12:04:36.204221+00:00"},{"alias_kind":"pith_short_12","alias_value":"3ZDXG3WWLTJR","created_at":"2026-07-05T12:04:36.204221+00:00"},{"alias_kind":"pith_short_16","alias_value":"3ZDXG3WWLTJRVGGW","created_at":"2026-07-05T12:04:36.204221+00:00"},{"alias_kind":"pith_short_8","alias_value":"3ZDXG3WW","created_at":"2026-07-05T12:04:36.204221+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":12,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.06928","citing_title":"VoxCPM2 Technical Report","ref_index":29,"is_internal_anchor":false},{"citing_arxiv_id":"2606.05858","citing_title":"ReverseEOL: Improving Training-free Text Embeddings via Text Reversal in Decoder-only LLMs","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2606.04511","citing_title":"SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference","ref_index":14,"is_internal_anchor":false},{"citing_arxiv_id":"2605.18753","citing_title":"DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention","ref_index":19,"is_internal_anchor":false},{"citing_arxiv_id":"2507.21166","citing_title":"The Ratchet Effect in Silico: How Interaction Drives Cumulative Intelligence in Large Language Models","ref_index":43,"is_internal_anchor":false},{"citing_arxiv_id":"2510.26692","citing_title":"Kimi Linear: An Expressive, Efficient Attention Architecture","ref_index":99,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27607","citing_title":"JaiTTS: A Thai Voice Cloning Model","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27393","citing_title":"MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction","ref_index":25,"is_internal_anchor":false},{"citing_arxiv_id":"2605.09544","citing_title":"TIDE-Bench: Task-Aware and Diagnostic Evaluation of Tool-Integrated Reasoning","ref_index":35,"is_internal_anchor":false},{"citing_arxiv_id":"2604.27607","citing_title":"JaiTTS: A Thai Voice Cloning Model","ref_index":12,"is_internal_anchor":false},{"citing_arxiv_id":"2604.10733","citing_title":"Too Nice to Tell the Truth: Quantifying Agreeableness-Driven Sycophancy in Role-Playing Language Models","ref_index":46,"is_internal_anchor":false},{"citing_arxiv_id":"2604.14922","citing_title":"LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning","ref_index":26,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU","json":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU.json","graph_json":"https://pith.science/api/pith-number/3ZDXG3WWLTJRVGGWPMP2XHXORU/graph.json","events_json":"https://pith.science/api/pith-number/3ZDXG3WWLTJRVGGWPMP2XHXORU/events.json","paper":"https://pith.science/paper/3ZDXG3WW"},"agent_actions":{"view_html":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU","download_json":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU.json","view_paper":"https://pith.science/paper/3ZDXG3WW","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2506.07900&json=true","fetch_graph":"https://pith.science/api/pith-number/3ZDXG3WWLTJRVGGWPMP2XHXORU/graph.json","fetch_events":"https://pith.science/api/pith-number/3ZDXG3WWLTJRVGGWPMP2XHXORU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU/action/storage_attestation","attest_author":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU/action/author_attestation","sign_citation":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU/action/citation_signature","submit_replication":"https://pith.science/pith/3ZDXG3WWLTJRVGGWPMP2XHXORU/action/replication_record"}},"created_at":"2026-07-05T12:04:36.204221+00:00","updated_at":"2026-07-05T12:04:36.204221+00:00"}