[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273475-105":59,"doc-detail-273475-en":135},{"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":128,"head_meta":130,"extra_data":132,"updated_unix":134},105,"en","palmyra-x6-technical-report-an-agentic-tool-use-model-post-trained-via-anchored-supervised-fine-tuning","Palmyra x6 Technical Report - An Agentic, Tool-Use Model Post-Trained via Anchored Supervised Fine-Tuning","","Palmyra x6 is a post-trained flagship open Mixture-of-Experts LLM designed for agentic task completion, including multi-step planning, tool/function calling, and long-horizon execution in tool-rich environments such as MCP suites, web search stacks, code-execution sandboxes, and retrieval-augmented pipelines. The report details a two-phase development process: systematic recipe exploration using SFT vs. ASFT with KL anchoring, followed by final training that changes only the base model and optimizer while keeping data and the anchored objective. Evaluations emphasize internal customer-like domains rather than public benchmarks, and the document covers architecture, data synthesis and filtering, objective and optimization, protocol, plus security and safety posture.",{"@graph":69,"@context":127},[70,84,106],{"@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/palmyra-x6-technical-report-an-agentic-tool-use-model-post-trained-via-anchored-supervised-fine-tuning/273475/",{"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/palmyra-x6-technical-report-an-agentic-tool-use-model-post-trained-via-anchored-supervised-fine-tuning/273475.png","ImageObject",300,407,{"name":92,"@type":93},"Emma Wilson","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-27","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",15,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119,123],{"name":110,"@type":111,"acceptedAnswer":112},"What capability is Palmyra x6 designed to improve?","Question",{"text":113,"@type":114},"It targets agentic task completion, including multi-step planning, tool/function calling, and long-horizon completion in tool-rich environments.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How did the authors develop the training recipe for Palmyra x6?",{"text":118,"@type":114},"They ran a matrix of twelve SFT/ASFT runs varying objectives, KL-anchor strength, learning rates, epoch counts, and six data recipes, then selected ASFT with KL weight K=0.1 over the primary mix.",{"name":120,"@type":111,"acceptedAnswer":121},"What changes were made in the final training phase?",{"text":122,"@type":114},"Final training kept the same data and ASFT objective while changing exactly two elements: the base model (a more advanced model) and the optimizer (Muon instead of Adam for specific weight matrices).",{"name":124,"@type":111,"acceptedAnswer":125},"Why does the report rely on internal benchmarks instead of public benchmarks as the main rubric?",{"text":126,"@type":114},"Public benchmarks are used as a sanity check, while internal datasets mimic enterprise customer domains and use cases, better reflecting production behavior.","https://schema.org",{"og:url":83,"og:type":129,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":131,"canonical":83},"index,follow",{"doc_id":133,"site_id":62},273475,1789471612,{"code":4,"msg":5,"data":136},{"doc_id":133,"user_id":137,"nickname":92,"user_avatar":138,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":139,"file_id":140,"file_url":141,"file_type":142,"file_size":143,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":144,"language":145,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":146,"faqs":147,"seo_title":148,"seo_description":67,"update_tm":134,"read_time":41,"preview_page_count":4},3848291630094,"https://eur-avatar.wpscdn.com/davatar_085a072bc5b1113ac321206ff7593b45","arXiv : 2608 . 16620v 3 [cs .CL] 8 Sep 2026  \nPalmyra x6 Technical Report An Agentic, Tool-Use Model Post-Trained via  \nAnchored Supervised Fine-Tuning  \nPeng Du* Kiran Kamble Rakshith Vasudev Zhizhuo Yang  \nRohith Nadimpally Arjun Krishna Waseem AlShikh  \nDaniel M. Bikel*  \nWriter AI Research, Writer, Inc.  \n{peng, kiran, rakshith, zhizhuo, rohith, arjun, waseem, [dan.bikel}@writer.com](dan.bikel}@writer.com)  \nAugust 12, 2026  \n1 Introduction  \nHistorically, in order to produce a large language model that was high-performing in many dimensions and excelled in a relatively small subset of interest, one would often take on the entire training pipeline, from pretraining, to continued pretraining/mid-training, to the various stages of post-training. Recently, however, it often suffices to post-train an existing high-performing LLM, provided that the training data is of high quality and the objective is carefully chosen to match the needs of the model’s intended applications. Palmyra x6 is a case in point. The primary capability we are interested in is agentic task completion: multi-step planning, tool/function calling, and long-horizon task completion within tool-rich environments such as MCP (Model Context Protocol) suites, web search stacks, code-execution sandboxes, and retrieval-augmented pipelines. After extensive testing, we chose a flagship open Mixture-of-Experts (MoE) base with 744B total and ∼40B active parameters per token. To it we apply a compact, high-quality corpus of synthetic agentic trajectories and an anchored fine-tuning objective [1] designed to add agentic skill without eroding the abilities the base model already has.  \nIn terms of task performance, Palmyra x6 delivers a huge improvement over the previous default model for Writer Agent, as shown by Figure 1. Comparisons against other frontier models available at the time of development are collected in Figure 2. However, we note that we use public benchmark performance merely as a sanity check and a guide, but not as the ultimate rubric. Rather, our focus was on developing internal datasets that were meant to mimic our enterprise customers’ domains and use cases, and closely track model performance on those, as they are a much better reflection of how the model will do in production than public benchmarks.  \n2 Overview  \nOur model development proceeded in two phases:  \n1. Recipe development. We ran a systematic matrix of twelve SFT and ASFT training runs, exploring objectives (plain SFT vs. ASFT), KL-anchor strengths, learning rates, epoch counts, and six data recipes. This phase established the primary candidate recipe: ASFT with KL weight K=0 .1 over the 12-dataset primary mix.  \n2. Final training and optimization. Palmyra x6 keeps the data and the ASFT objective and changes exactly two things: the base model (employing a more advanced base model) and the optimizer (Muon [2] instead of Adam for the 2-D weight matrices) .  \n*Corresponding authors: dan .bikel@writer .com and peng@writer .com.  \nFigure 1: Per-benchmark comparison of Palmyra x6 against the previous default model for Writer Agent on six evaluation benchmarks, ordered by margin. Bars show mean score (higher is better); error bars on Palmyra x6 denote one standard error. Palmyra x6 leads on every benchmark, with the largest gains on MCP-Atlas (+0 .320), FinanceBench (+0 .305), and IFBench (+0 .304) .  \nWe were careful to isolate variables wherever possible: we had one data recipe and one objective with the two controlled changes listed above. The ultimate effect is salutary: any capability delta between the initial model development candidates and Palmyra x6 is attributable to the new base model and optimizer, and not to confounding changes in data or loss.  \nThe remainder of this report proceeds as follows. Section 3 describes the architecture we inherit from the base model; Section 4, the data we synthesize and the filters every trajectory must survive; Section 5, the objective, the optimizer","cbCaieI8KRhi2LU3","https://ap.wps.com/l/cbCaieI8KRhi2LU3","pdf",644578,12,"English","# Introduction\n# Overview\n# Model Architecture\n# Training Data\n# Objective and Optimization\n# Evaluation Protocol and Results\n# Security and Safety Posture\n# Limitations\n# Conclusion","[{\"question\":\"What capability is Palmyra x6 designed to improve?\",\"answer\":\"It targets agentic task completion, including multi-step planning, tool/function calling, and long-horizon completion in tool-rich environments.\"},{\"question\":\"How did the authors develop the training recipe for Palmyra x6?\",\"answer\":\"They ran a matrix of twelve SFT/ASFT runs varying objectives, KL-anchor strength, learning rates, epoch counts, and six data recipes, then selected ASFT with KL weight K=0.1 over the primary mix.\"},{\"question\":\"What changes were made in the final training phase?\",\"answer\":\"Final training kept the same data and ASFT objective while changing exactly two elements: the base model (a more advanced model) and the optimizer (Muon instead of Adam for specific weight matrices).\"},{\"question\":\"Why does the report rely on internal benchmarks instead of public benchmarks as the main rubric?\",\"answer\":\"Public benchmarks are used as a sanity check, while internal datasets mimic enterprise customer domains and use cases, better reflecting production behavior.\"}]","Palmyra x6 Technical Report - An Agentic, Tool-Use Model Post-Trained via Anchored Supervised Fine-Tuning | PDF"]