[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-515540-105":59,"doc-detail-515540-en":134},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":127,"head_meta":129,"extra_data":131,"updated_unix":133},105,"en","hermes-4-technical-report","HERMES 4 TECHNICAL REPORT","","Hermes 4 introduces a family of hybrid reasoning models that merge structured, multi-turn reasoning with strong instruction-following capabilities. The report details data curation, synthesis, training, and evaluation challenges, then presents scalable solutions for each stage. Comprehensive assessments span mathematical reasoning, coding, knowledge, comprehension, and alignment, reporting both quantitative results and qualitative behavioral analysis. Model weights are published publicly to support open research.",{"@graph":69,"@context":126},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/hermes-4-technical-report/515540/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/hermes-4-technical-report/515540.png","ImageObject",300,407,{"name":92,"@type":93},"วิน","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-15","2026-10-06",true,{"@type":102,"interactionType":103,"userInteractionCount":34},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118,122],{"name":109,"@type":110,"acceptedAnswer":111},"What is Hermes 4 and what capabilities does it target?","Question",{"text":112,"@type":113},"Hermes 4 is a family of hybrid reasoning models combining structured multi-turn reasoning with broad instruction-following ability. It is designed to enhance reasoning while retaining general knowledge and instruction competence.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"How is the Hermes 4 dataset constructed?",{"text":117,"@type":113},"The dataset is largely newly synthesized, totaling about 5 million samples and 19 billion tokens. It uses a hybrid approach with reasoning-focused and non-reasoning instruction data, and retains a portion of the Hermes 3 dataset for continuity.",{"name":119,"@type":110,"acceptedAnswer":120},"What is DataForge in the post-training pipeline?",{"text":121,"@type":113},"DataForge is a graph-based synthetic data generator that creates conversational data across many task types. It uses a directed acyclic graph where each node specifies struct-to-struct mappings via a PDDL action interface.",{"name":123,"@type":110,"acceptedAnswer":124},"How are Hermes 4 models evaluated?",{"text":125,"@type":113},"Evaluation covers mathematical reasoning, coding, knowledge, comprehension, and alignment benchmarks. The report includes both quantitative performance metrics and qualitative behavioral analysis, and provides publicly released model weights for reproducibility.","https://schema.org",{"og:url":83,"og:type":128,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":130,"canonical":83},"index,follow",{"doc_id":132,"site_id":62},515540,1791446808,{"code":4,"msg":5,"data":135},{"doc_id":132,"user_id":136,"nickname":92,"user_avatar":137,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":138,"file_id":139,"file_url":140,"file_type":141,"file_size":142,"view_count":34,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":36,"language":143,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":144,"faqs":145,"seo_title":146,"seo_description":67,"update_tm":147,"read_time":148,"preview_page_count":19,"preview_urls":149},2336475104736,"https://ap-avatar.wpscdn.com/avatar/22000c4c5e0e5b17e70?x-image-process=image/resize,m_fixed,w_180,h_180&k=1786591360781797222","HERMES 4 TECHNICAL REPORT  \narXiv :2508 . 18255v2 [ cs .AI] 2 Sep 2025  \nRyan Teknium*  \nX: @Teknium1  \nRoger Jin  \nX: @rogershijin  \nJai Suphavadeeprasit  \nX: @JSupa15  \nDakota Mahan  \nX: @dmayhem93  \nJeffrey Quesnelle  \nX: @theemozilla  \nJoe Li  \nX: @JoeLi5050  \nChen Guang  \nX: @nullvaluetensor  \nShannon Sands  \nX: @max_paperclips  \nKaran Malhotra  \nX: @karan4d  \nABSTRACT  \nWe present Hermes 4, a family of hybrid reasoning models that combine structured, multi-turn reasoning with broad instruction-following ability. We describe the challenges encountered during data curation, synthesis, training, and evaluation, and outline the solutions employed to address these challenges at scale.  \nWe comprehensively evaluate across mathematical reasoning, coding, knowledge, comprehension, and alignment benchmarks, and we report both quantitative performance and qualitative behavioral analysis. To support open research, all model weights are published publicly at [https://huggingface.co/NousResearch](https://huggingface.co/NousResearch2)[2](https://huggingface.co/NousResearch2) .  \n1 Introduction  \nLLMs, at this point, need no introduction. Their effectiveness at approximating aspects of human cognition despite comparatively simple training objectives has made them a subject of both academic investigation and practical deployment. In recent years, a growing body of work has explored inference-time scaling mechanisms (often described as “reasoner\" models) which dynamically adjust the amount of computation devoted to solving a task [35, 15] .  \nInitially, reasoner models were available only from proprietary providers. More recently, a number of open-weight reasoning systems have been released, enabling the community to study, evaluate, and improve upon these methods in an open setting [57, 15, 48] .  \nIn this report we introduce Hermes 4, a family of neutrally-aligned generalist models trained to integrate self-reflective reasoning with broad instructional competence. The contributions of this work are threefold:  \n• A data synthesis and curation strategy that produces a large-scale hybrid dataset consisting of both reasoning-focused and general purpose instruction examples (Section2);  \n• A training methodology that incorporates loss-masking, length-control fine-tuning, and efficient packing strategies for large-scale heterogeneous data (Section 3);  \n∗ [Correspondence to](Correspondence to teknium@nousresearch.com)[ teknium@nousresearch.com](Correspondence to teknium@nousresearch.com).  \n2[https://huggingface.co/collections/NousResearch/hermes-4-collection-68a731bfd452e20816725728](https://huggingface.co/collections/NousResearch/hermes-4-collection-68a731bfd452e20816725728)  \n• A comprehensive evaluation across reasoning, code, knowledge, alignment, and qualitative benchmarks (Section 4, 5) .  \nTogether, these contributions demonstrate how open-weight reasoning models can be effectively trained and rigorously evaluated, yielding models that are comparable to frontier systems while remaining transparent and reproducible.  \n2 Post-training Data  \nThe Hermes 4 dataset consists primarily of newly synthesized reasoning and non-reasoning data, totaling approximately 5 million samples and 19 billion tokens. It was designed to imbue the model with advanced reasoning capabilities while maintaining a broad base of general knowledge and instruction-following proficiency. The data strategy focused on a hybrid approach, combining a substantial volume of reasoning-focused data with a diverse set of non-reasoning instructions. The dataset is a composite of 3.5 million reasoning samples and 1.6 million non-reasoning samples. A significant portion of the Hermes 3 [51] dataset3 was retained to ensure continuity in the model’s capabilities. The reasoning samples were intentionally token-heavy, with an average of five times more tokens per sample than their non-reasoning counterparts, accommodating thinking traces up to 16 thousand tokens long.  \n2.1 DataFo","cbCaipOibKKP6e71","https://ap.wps.com/l/cbCaipOibKKP6e71","pdf",1656768,"English","# Introduction\n# Post-training Data\n## DataForge\n### Pre-training seed data cleaning and deduplication\n### An example DataForge graph","[{\"question\":\"What is Hermes 4 and what capabilities does it target?\",\"answer\":\"Hermes 4 is a family of hybrid reasoning models combining structured multi-turn reasoning with broad instruction-following ability. It is designed to enhance reasoning while retaining general knowledge and instruction competence.\"},{\"question\":\"How is the Hermes 4 dataset constructed?\",\"answer\":\"The dataset is largely newly synthesized, totaling about 5 million samples and 19 billion tokens. It uses a hybrid approach with reasoning-focused and non-reasoning instruction data, and retains a portion of the Hermes 3 dataset for continuity.\"},{\"question\":\"What is DataForge in the post-training pipeline?\",\"answer\":\"DataForge is a graph-based synthetic data generator that creates conversational data across many task types. It uses a directed acyclic graph where each node specifies struct-to-struct mappings via a PDDL action interface.\"},{\"question\":\"How are Hermes 4 models evaluated?\",\"answer\":\"Evaluation covers mathematical reasoning, coding, knowledge, comprehension, and alignment benchmarks. The report includes both quantitative performance metrics and qualitative behavioral analysis, and provides publicly released model weights for reproducibility.\"}]","HERMES 4 TECHNICAL REPORT | PDF",1791317458,101,[150,154,158,162],{"page_no":8,"url":151,"width":152,"height":153},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/51144601173e8a8266f6b3cc53e902df.webp",983,1272,{"page_no":14,"url":155,"width":156,"height":157},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/4cb53c9b6179d8e6bc4f96cda6c077e6.webp",982,1270,{"page_no":81,"url":159,"width":160,"height":161},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/4569756fb8f82e9ef11f804bfaf7af81.webp",981,1269,{"page_no":19,"url":163,"width":164,"height":165},"https://res-academy.cache.wpscdn.com/upload/kdocshare/2026-11-17/8c56a9ab41e12c9a7eb0b9266d3a17fc.webp",980,1268]