[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82330-en":3,"doc-seo-82330-105":29,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},82330,1374391974564,"Clementine","https://ap-avatar.wpscdn.com/avatar/14000253aa45c000a9e?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779874745381141002",8,"Research & Report","Mach Mind 4 Flash Technical Report","Mach-Mind-4-Flash is a 35B-parameter Mixture-of-Experts (MoE) agentic model with 3B activated parameters that matches or exceeds the performance of 100B-class models without scaling pre-training compute, relying on post-training optimization only. The system uses scalable agentic interaction environments for large-scale reinforcement learning and a three-stage pipeline: unified RL/OPD infrastructure, multi-track expert training fused by Multi-Teacher On-Policy Distillation, and Hybrid Median-length Policy Optimization for token-efficient reasoning compression with minimal accuracy loss.","arXiv :2607 .09375v 1 [ cs .LG] 10 Jul 2026  \nMach-Mind-4-Flash Technical Report  \nFoundation Model, Li Auto Inc.  \nAbstract  \nWe present Mach-Mind-4-Flash, a 35B-parameter Mixture-of-Experts (MoE) agentic model with 3B activated parameters. Through post-training optimization alone without scaling pre-training compute, the model achieves performance on par with or surpassing that of 100B-parameter-class models. By introducing scalable agentic interaction environments for large-scale reinforcement learning, the model attains significant performance gains on real-world application tasks. Our pipeline comprises three stages: (1) a unified RL/OPD training infrastructure with dynamic multi-teacher scheduling and operator-level acceleration, delivering 17% end-toend training speedup; (2) multiple domain-specific RL experts trained in parallel across Reasoning, General, and Agent tracks, then fused into a single generalist via Multi-Teacher On-Policy Distillation (MOPD)—a routed reverse-KL objective that eliminates the see-saw degradation of mixed-reward RL; (3) Hybrid Medianlength Policy Optimization (HMPO), a single-stage token-efficiency method that compresses reasoning chains by 19–46% with ≤0.7 percentage-point accuracy loss. Mach-Mind-4-Flash scores 92 .70 on AIME’26, 82 . 82 on IFBench, 80 .74 on Behavioral-SafetyBench, 75.80 on BFCL-v4, 72.31 on BrowseComp-zh, and 84.20 on ClawBench—leading or matching models with 10–30× its activated size at a fraction of the inference cost.  \n Mach-Mind-4-Flash  \n MiMo-V2-Flash-309B-A15B  \nQwen3.5-35B-A3BQwen3.5-122B-A10B  \n GLM-4.7-Flash-30B-A3B  \n Step-3.5-Flash-196B-A11B  \nNemotron-3-120B-A12B Kimi-K2.5-1T-A32B  \nFigure 1: Mach-Mind-4-Flash matches or exceeds much larger models across diverse capability axes. With only 3B activated parameters, Mach-Mind-4-Flash leads on IFBench, BehavioralSafetyBench, and BrowseComp-zh, while remaining competitive on reasoning, tool use, and agentic coding against models with 3–30× its activated size.  \nContents  \n1 Introduction 3  \n2 Related Work 3  \n2.1 Agentic Foundation Models and Benchmarks .................... 4  \n2.2 Agentic Post-training and Reinforcement Learning ................. 4  \n2.3 Efficient and Safe Agent Deployment ........................ 4  \n3 Infra 5  \n3.1 Unified OPD Training Paradigm Based on the RL Framework ........... 5  \n3.2 Dynamic Multi-Teacher Scalable Architecture .................... 6  \n3.3 Training Acceleration and Extreme Optimization .................. 7  \n4 Post-Training 8  \n4.1 Supervised Fine-Tuning ................................ 8  \n4.2 Reasoning and General RL .............................. 9  \n4.3 Safety RL ....................................... 11  \n4.4 Tool-Use RL ...................................... 11  \n4.5 DeepSearch RL .................................... 13  \n4.6 Code Agent RL .................................... 14  \n4.7 Claw Agent RL .................................... 15  \n4.8 Multi-Teacher On-Policy Distillation (MOPD) .................... 17  \n4.9 Hybrid Median-length Policy Optimization (HMPO) ................ 18  \n5 Experimental Results 20  \n5.1 Overall Results .................................... 20  \n5.2 Effect of Expert Training and MOPD Fusion ..................... 21  \n5.3 Token efficient (HMPO) ............................... 22  \n6 Limitation and Future Work 22  \n7 Contributions 23  \nAppendix 31  \nA Acceleration of the Infra operator 31  \nB Derivation of the MOPD Objective 31  \nC Ablations on the Reasoning Expert in MOPD 32  \n1 Introduction  \nScaling language models has been the dominant recipe for capability gains, yet the associated inference cost makes trillion-parameter models impractical for latency-sensitive deployment. An emerging alternative is to scale the post-training pipeline, pushing a compact base model toward frontier performance through reinforcement learning, expert fusion, and inference-time efficiency optimization rather than scaling pre-training compute alone. Mach-M","cbCaidw8kf9kuIfp","https://ap.wps.com/l/cbCaidw8kf9kuIfp","pdf",9693513,1,33,"English","en",105,"# Introduction\n# Related Work\n## Agentic Foundation Models and Benchmarks\n## Agentic Post-training and Reinforcement Learning\n## Efficient and Safe Agent Deployment\n# Infra\n## Unified OPD Training Paradigm Based on the RL Framework\n## Dynamic Multi-Teacher Scalable Architecture\n## Training Acceleration and Extreme Optimization\n# Post-Training\n## Supervised Fine-Tuning\n## Reasoning and General RL\n## Safety RL\n## Tool-Use RL\n## DeepSearch RL\n## Code Agent RL\n## Claw Agent RL\n## Multi-Teacher On-Policy Distillation (MOPD)\n## Hybrid Median-length Policy Optimization (HMPO)\n# Experimental Results\n## Overall Results\n## Effect of Expert Training and MOPD Fusion\n## Token Efficient (HMPO)\n# Limitation and Future Work\n# Contributions\n# Appendix","[{\"question\":\"What is Mach-Mind-4-Flash and what are its key model characteristics?\",\"answer\":\"Mach-Mind-4-Flash is a 35B-parameter Mixture-of-Experts agentic model with 3B activated parameters. It is designed to reach performance comparable to or better than 100B-class models via post-training optimization rather than scaling pre-training compute.\"},{\"question\":\"How does the training pipeline of Mach-Mind-4-Flash work?\",\"answer\":\"The pipeline has three stages: a unified RL/OPD training infrastructure with dynamic multi-teacher scheduling and operator-level acceleration; parallel training of domain-specific RL experts and fusion through Multi-Teacher On-Policy Distillation (MOPD); and a token-efficiency method, Hybrid Median-length Policy Optimization (HMPO), to compress reasoning chains with limited accuracy impact.\"},{\"question\":\"What problem does MOPD address when combining multiple RL experts?\",\"answer\":\"MOPD introduces a routed reverse-KL objective to eliminate the see-saw degradation commonly caused by mixed-reward RL when training across heterogeneous reward settings. It consolidates specialized experts into a single generalist by supervising the student on rollouts aligned to each teacher.\"}]",1784179675,83,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"mach-mind-4-flash-technical-report","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/mach-mind-4-flash-technical-report/82330/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What is Mach-Mind-4-Flash and what are its key model characteristics?","Question",{"text":75,"@type":76},"Mach-Mind-4-Flash is a 35B-parameter Mixture-of-Experts agentic model with 3B activated parameters. It is designed to reach performance comparable to or better than 100B-class models via post-training optimization rather than scaling pre-training compute.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the training pipeline of Mach-Mind-4-Flash work?",{"text":80,"@type":76},"The pipeline has three stages: a unified RL/OPD training infrastructure with dynamic multi-teacher scheduling and operator-level acceleration; parallel training of domain-specific RL experts and fusion through Multi-Teacher On-Policy Distillation (MOPD); and a token-efficiency method, Hybrid Median-length Policy Optimization (HMPO), to compress reasoning chains with limited accuracy impact.",{"name":82,"@type":73,"acceptedAnswer":83},"What problem does MOPD address when combining multiple RL experts?",{"text":84,"@type":76},"MOPD introduces a routed reverse-KL objective to eliminate the see-saw degradation commonly caused by mixed-reward RL when training across heterogeneous reward settings. It consolidates specialized experts into a single generalist by supervising the student on rollouts aligned to each teacher.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]