[{"data":1,"prerenderedAt":2602},["ShallowReactive",2],{"sanity-7iXP3sOrSyrwpMoyaopqutrI_r2a7XNibHtOh578tGQ":3,"sanity-TVCNm1l9XOkpZpU2Ys9lHUizSiqOFMfqBK_1-dQkuzk":219,"sanity-a19VNodgPylkcMhsC0-wFY8WxK4WWZeI5txIHkZt2Sk":419},{"data":4,"sourceMap":139},{"_createdAt":5,"_id":6,"_rev":7,"_system":8,"_type":6,"_updatedAt":11,"defaultTitle":12,"footerMenu1":13,"footerMenu2":29,"footerMenu3":44,"gitHubUrl":53,"huggingFaceUrl":57,"linkedInUrl":49,"mainNavigation":59,"ogDescription":87,"ogImage":88,"ogImageUrl":93,"showSiteNotice":94,"siteNotice":95},"2026-01-21T14:58:26Z","siteSettingsSchema","QKcLXhFVWa14MNDoKwaAu5",{"base":9},{"id":6,"rev":10},"O3kmd93YEHgcHWK9rNfBS8","2026-07-10T07:17:33Z","kaiko.ai",{"links":14,"title":28},[15,20,24],{"_key":16,"_type":17,"label":18,"url":19},"95cf27020263","navigationItem","Clinical AI Workspace","/clinical-ai-workspace",{"_key":21,"_type":17,"label":22,"url":23},"7b5bdfebb176","Research","/research",{"_key":25,"_type":17,"label":26,"url":27},"d311ee16a390","Insights Hub","/insights-hub","What we do",{"links":30,"title":43},[31,35,39],{"_key":32,"_type":17,"label":33,"url":34},"44a37ad9151d","About us","/about",{"_key":36,"_type":17,"label":37,"url":38},"193660568291","Careers","https://jobs.kaiko.ai/",{"_key":40,"_type":17,"label":41,"url":42},"aeb117deb742","Trust Center","https://trust.kaiko.ai/","Company",{"links":45,"title":58},[46,50,54],{"_key":47,"_type":17,"label":48,"url":49},"40c6bf52fdd1","LinkedIn","https://www.linkedin.com/company/kaiko-ai/",{"_key":51,"_type":17,"label":52,"url":53},"f79c6a57ca2f","GitHub","https://github.com/kaiko-ai",{"_key":55,"_type":17,"label":56,"url":57},"7ff3680a1f90","Hugging Face","https://huggingface.co/kaiko-ai","Community",[60,62,64],{"_key":61,"_type":17,"label":18,"url":19},"601985fd4221",{"_key":63,"_type":17,"label":22,"url":23},"a10cc775319c",{"_key":65,"_type":66,"children":67,"label":86},"3c77dba0a31e","navigationGroup",[68,71,73,77,79,82],{"_key":69,"_type":17,"label":70,"url":34},"4d07fcce85a7","About",{"_key":72,"_type":17,"label":26,"url":27},"66b87567d450",{"_key":74,"_type":17,"label":75,"url":76},"bf5d3c3b0672","Events","https://kaiko.ai/insights-hub?categories=event",{"_key":78,"_type":17,"label":37,"url":38},"b34eb970c9d9",{"_key":80,"_type":17,"label":81,"url":42},"0179b3449f54","Security & Compliance",{"_key":83,"_type":17,"label":84,"url":85},"9f1fac3acbae","Contact Us","mailto:info@kaiko.ai","Resources","Waypoint helps solo builders shape rough ideas, refine implementation prompts, preserve project decisions, and scan repositories for common readiness risks before shipping to users.",{"_type":89,"asset":90},"image",{"_ref":91,"_type":92},"image-d9426451401bfd3574519485220b90bce4571422-1200x630-jpg","reference","https://cdn.sanity.io/images/a5kt9um3/production/d9426451401bfd3574519485220b90bce4571422-1200x630.jpg?w=1200&h=630&fit=crop",false,{"badges":96,"link":125,"linkText":126,"message":127},[97,118],{"_key":98,"color":99,"label":117},"70498638b5b3",{"_type":100,"alpha":101,"hex":102,"hsl":103,"hsv":108,"rgb":112},"color",1,"#a8e2df",{"_type":104,"a":101,"h":105,"l":106,"s":107},"hslaColor",176.89655172413794,0.7725490196078431,0.4999999999999999,{"_type":109,"a":101,"h":105,"s":110,"v":111},"hsvaColor",0.2566371681415929,0.8862745098039215,{"_type":113,"a":101,"b":114,"g":115,"r":116},"rgbaColor",223,226,168,"Online event",{"_key":119,"color":120,"label":124},"b238b246731c",{"_type":100,"alpha":101,"hex":102,"hsl":121,"hsv":122,"rgb":123},{"_type":104,"a":101,"h":105,"l":106,"s":107},{"_type":109,"a":101,"h":105,"s":110,"v":111},{"_type":113,"a":101,"b":114,"g":115,"r":116},"Rewatch now","/insights-hub/kaiko-launch-watch","Watch here",[128],{"_key":129,"_type":130,"children":131,"markDefs":137,"style":138},"515be88d5691","block",[132],{"_key":133,"_type":134,"marks":135,"text":136},"217241b6c7c6","span",[],"Pilot limbo: why most clinical AI never scales, and how some hospitals break the cycle",[],"normal",{"documents":140,"paths":144,"mappings":163},[141,143],{"_id":91,"_type":142},"sanity.imageAsset",{"_id":6,"_type":6},[145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162],"$['_createdAt']","$['_id']","$['_rev']","$['_system']","$['_type']","$['_updatedAt']","$['defaultTitle']","$['footerMenu1']","$['footerMenu2']","$['footerMenu3']","$['gitHubUrl']","$['huggingFaceUrl']","$['linkedInUrl']","$['mainNavigation']","$['ogDescription']","$['ogImage']","$['siteNotice']","$['url']",{"$['_createdAt']":164,"$['_id']":169,"$['_rev']":171,"$['_system']":174,"$['_type']":177,"$['_updatedAt']":180,"$['defaultTitle']":183,"$['footerMenu1']":186,"$['footerMenu2']":189,"$['footerMenu3']":192,"$['gitHubUrl']":195,"$['huggingFaceUrl']":198,"$['linkedInUrl']":201,"$['mainNavigation']":204,"$['ogDescription']":207,"$['ogImage']":210,"$['ogImageUrl']":213,"$['siteNotice']":216},{"source":165,"type":168},{"document":101,"path":166,"type":167},0,"documentValue","value",{"source":170,"type":168},{"document":101,"path":101,"type":167},{"source":172,"type":168},{"document":101,"path":173,"type":167},2,{"source":175,"type":168},{"document":101,"path":176,"type":167},3,{"source":178,"type":168},{"document":101,"path":179,"type":167},4,{"source":181,"type":168},{"document":101,"path":182,"type":167},5,{"source":184,"type":168},{"document":101,"path":185,"type":167},6,{"source":187,"type":168},{"document":101,"path":188,"type":167},7,{"source":190,"type":168},{"document":101,"path":191,"type":167},8,{"source":193,"type":168},{"document":101,"path":194,"type":167},9,{"source":196,"type":168},{"document":101,"path":197,"type":167},10,{"source":199,"type":168},{"document":101,"path":200,"type":167},11,{"source":202,"type":168},{"document":101,"path":203,"type":167},12,{"source":205,"type":168},{"document":101,"path":206,"type":167},13,{"source":208,"type":168},{"document":101,"path":209,"type":167},14,{"source":211,"type":168},{"document":101,"path":212,"type":167},15,{"source":214,"type":168},{"document":166,"path":215,"type":167},17,{"source":217,"type":168},{"document":101,"path":218,"type":167},16,{"data":220,"sourceMap":369},{"_createdAt":221,"_id":222,"_rev":223,"_system":224,"_type":222,"_updatedAt":227,"getInTouch":228,"insights":235,"intro":239,"kaiko":244,"metaDescription":87,"metaTitle":265,"partners":266,"press":311,"research":346,"security":351},"2026-02-03T11:33:56Z","homePage","fEgPm0GnAQb0S9IeucSPZL",{"base":225},{"id":222,"rev":226},"xGh7Bt6z4iPTi5vmNY4hts","2026-06-25T12:22:06Z",{"heading":229,"newsletterSubtitle":232,"newsletterTitle":233,"text":234},{"blueText":230,"gradientText":231},"Bring Waypoint into","your next build","Join our newsletter for curated insights at the intersection of AI and healthcare.","Get the latest","Whether you are shaping an idea or preparing to ship, Waypoint keeps decisions connected through the journey.",{"categoryIds":236,"heading":237,"subHeading":238},null,"Latest news & research","Stay updated with AI innovations in healthcare.",{"eyebrow":240,"heading":241,"subHeading":242,"text":243},"Build\ndeliberately","Turn your rough ideas into\nbuild-ready plans with confidence.","Waypoint is the intelligence layer around every vibe-coded project: sharpen each prompt, record every decision scan the repository before you ship your work to real users.","Waypoint does not write or host your application. Instead, it strengthens the work around your coding agent: shaping the first brief, sharpening feature prompts, preserving why decisions changed, and checking the repository before release. You keep building in Claude Code, Cursor, or whichever harness already best fits your workflow.",{"eyebrow":245,"heading":246,"sections":247,"text":264},"Your guided build companion","Confident builds start with a clear route.",[248,252,256,260],{"_key":249,"text":250,"title":251},"c88f7d7dc339","Focused questions turn rough ideas into structured, reviewable briefs for builders.","Turn ideas into\nclear briefs",{"_key":253,"text":254,"title":255},"0cea88e9b15a","Targeted refinement produces precise prompts for your preferred coding agent.","Sharpen\nfeature prompts",{"_key":257,"text":258,"title":259},"dc68c963e5de","Structured logs preserve what changed, why it changed, and what happens next clearly.","Preserve every\nbuild decision",{"_key":261,"text":262,"title":263},"fe7ea1462636","Read-only repository scans reveal common risks without executing or retaining your code.","Scan before\nyou ship","Plans shaped, prompts sharpened, and projects checked before users ever arrive.","Waypoint guides confident software builders",{"eyebrow":267,"heading":268,"quotesSection":271},"Builders",{"blueText":269,"gradientText":270},"Designed around","builder momentum",{"quotes":272,"text":309,"title":310},[273,283,292,301],{"_key":274,"_type":275,"author":276,"company":277,"photo":278,"role":281,"text":282},"79b5dd0d2334","quote","Alexander van Someren","LUMC",{"_type":89,"asset":279},{"_ref":280,"_type":92},"image-27b6040da02080f8213aab785cb3ecefbecaa71b-299x299-jpg","AI Product Manager","Waypoint turns the messy beginning of a software project into a dependable route you can trust.\n\nFocused questions expose assumptions, an approved brief defines the MVP, and a reusable master prompt gives your coding agent enough context to begin building without carrying an entire planning conversation forward.",{"_key":284,"_type":275,"author":285,"company":286,"photo":287,"role":290,"text":291},"3c202d6e0e8a","Luc Karssemakers","Antoni van Leeuwenhoek Hospital",{"_type":89,"asset":288},{"_ref":289,"_type":92},"image-6836bead06bf3458260f386e2c6cd8e5d8be62c4-600x604-png","Head and Neck surgeon, \nDepartment of Head and Neck surgery","When the next feature is still vague, Waypoint helps turn intent into an implementation-ready prompt. It asks targeted questions, captures the important constraints, and produces a complete practical artifact for Claude Code or Cursor.\n\nThe result is less repeated context, fewer improvised decisions, and a clearer handoff every time you return to building, while the running build log preserves what changed, why it changed, and what should happen next.",{"_key":293,"_type":275,"author":294,"company":295,"photo":296,"role":299,"text":300},"0673a764a756","Dr. Kevin Groot Lipman","Antoni van Leeuwenhoek",{"_type":89,"asset":297},{"_ref":298,"_type":92},"image-071229b73508699778c68add9a3dee8f24b6cca0-3072x4096-jpg","Clinical AI Implementation Lead","Before release, Waypoint connects to your GitHub repository with read-only access and checks the latest commit for common risks. Source files remain transient, code is never executed, and only bounded, redacted findings persist.\n\nYou can triage what matters, compare scans, and see whether critical issues still block readiness, without turning a lightweight sanity check into a false claim of complete security.",{"_key":302,"_type":275,"author":303,"company":295,"photo":304,"role":307,"text":308},"eecccd01b367","Marene Kroon",{"_type":89,"asset":305},{"_ref":306,"_type":92},"image-7266fb82102b62bd85f6831c212962aa503af814-508x536-png","Nurse Practitioner","Waypoint keeps the work around your code connected from beginning to release. Briefs explain what you are building, enhanced prompts guide the next feature, build logs preserve why decisions changed, and repository scans reveal the risks worth fixing before real users arrive with confidence.","Waypoint keeps each phase connected, so your approved brief, sharper prompts, build decisions, and readiness findings stay organized together.","Built around your whole project",{"eyebrow":312,"heading":313,"logos":314},"Press","In the media",[315,322,328,334,340],{"_key":316,"_type":317,"logo":318,"url":321},"138f6f224b2c","logo",{"_type":89,"asset":319},{"_ref":320,"_type":92},"image-06e87ce731981c9ecbc23195194920f952260de4-203x15-svg","https://www.business-reporter.co.uk/management/the-ai-that-helps-doctors-see-cancers-complete-picture",{"_key":323,"_type":317,"logo":324,"url":327},"be6a97aa4114",{"_type":89,"asset":325},{"_ref":326,"_type":92},"image-7e20caaf8d01e7a96f232c959c958cccf8c825b2-225x45-svg","https://www.wired.com/sponsored/story/kaiko/",{"_key":329,"_type":317,"logo":330,"url":333},"d31f53f8e70f",{"_type":89,"asset":331},{"_ref":332,"_type":92},"image-495fbf2c3a25bf6e096c63cd004a869f37e9a38c-160x32-svg","https://www.tagesanzeiger.ch/weniger-papierkram-hoehere-qualitaet-jetzt-kommt-die-ki-aerztin-870977120920",{"_key":335,"_type":317,"logo":336,"url":339},"80857a8bc683",{"_type":89,"asset":337},{"_ref":338,"_type":92},"image-8517092e21b684e25c3b37a44d1b7d6727f437ae-125x22-svg","https://www.netzwoche.ch/news/2025-09-22/medizin-wird-praeziser-kein-zusammenhang-bleibt-unentdeckt?secure_access_token=r56ym8dy8g7nzgnj9sez7",{"_key":341,"_type":317,"logo":342,"url":345},"d3ed070adef9",{"_type":89,"asset":343},{"_ref":344,"_type":92},"image-bd6b703a746eb266c93f87caeae34afe14abfa82-207x40-svg","https://www.fuw.ch/robert-berke-kaiko-ai-seit-chatgpt-reicht-ein-satz-um-unser-produkt-zu-erklaeren-238508461486",{"eyebrow":22,"heading":347,"text":350},{"blueText":348,"gradientText":349},"The technology","behind it","We develop and study general-purpose AI systems designed for clinical environments: models that can reason over complex health data and assist care teams over multiple medical domains in real time.\n\nOur work spans foundation-model training, multimodal understanding, reasoning, safety and regulatory-grade model evaluation.",{"cta":352,"heading":354,"items":355},{"buttonText":41,"label":353,"url":42},"Full detail in our","Confidence before shipping.",[356,361,365],{"_key":357,"_type":358,"body":359,"title":360},"sec-certified","item","Focused questions turn rough ideas into structured briefs you can approve.","Briefs, approved.",{"_key":362,"_type":358,"body":363,"title":364},"sec-data","Only your authored product context reaches the AI refinement workflow.","Your reasoning stays yours.",{"_key":366,"_type":358,"body":367,"title":368},"sec-human","Waypoint scans repository archives transiently, never executes code, and stores only bounded, redacted findings that help you release with confidence.","Your repository stays private.",{"documents":370,"paths":372,"mappings":384},[371],{"_id":222,"_type":222},[145,146,147,148,149,150,373,374,375,376,377,378,379,380,381,382,383],"$['getInTouch']","$['insights']['heading']","$['insights']['subHeading']","$['intro']","$['kaiko']","$['metaDescription']","$['metaTitle']","$['partners']","$['press']","$['research']","$['security']",{"$['_createdAt']":385,"$['_id']":387,"$['_rev']":389,"$['_system']":391,"$['_type']":393,"$['_updatedAt']":395,"$['getInTouch']":397,"$['insights']['heading']":399,"$['insights']['subHeading']":401,"$['intro']":403,"$['kaiko']":405,"$['metaDescription']":407,"$['metaTitle']":409,"$['partners']":411,"$['press']":413,"$['research']":415,"$['security']":417},{"source":386,"type":168},{"document":166,"path":166,"type":167},{"source":388,"type":168},{"document":166,"path":101,"type":167},{"source":390,"type":168},{"document":166,"path":173,"type":167},{"source":392,"type":168},{"document":166,"path":176,"type":167},{"source":394,"type":168},{"document":166,"path":179,"type":167},{"source":396,"type":168},{"document":166,"path":182,"type":167},{"source":398,"type":168},{"document":166,"path":185,"type":167},{"source":400,"type":168},{"document":166,"path":188,"type":167},{"source":402,"type":168},{"document":166,"path":191,"type":167},{"source":404,"type":168},{"document":166,"path":194,"type":167},{"source":406,"type":168},{"document":166,"path":197,"type":167},{"source":408,"type":168},{"document":166,"path":200,"type":167},{"source":410,"type":168},{"document":166,"path":203,"type":167},{"source":412,"type":168},{"document":166,"path":206,"type":167},{"source":414,"type":168},{"document":166,"path":209,"type":167},{"source":416,"type":168},{"document":166,"path":212,"type":167},{"source":418,"type":168},{"document":166,"path":218,"type":167},{"data":420,"sourceMap":2453},[421,1103,1564,1748],{"_createdAt":422,"_id":423,"_rev":424,"_system":425,"_type":428,"_updatedAt":429,"area":430,"authors":432,"categories":436,"content":440,"date":1078,"featuredImage":1079,"isHidden":94,"relatedPosts":1082,"slug":1086,"subHeading":1089,"title":875,"topics":1090},"2026-07-09T16:13:00Z","5a60d1e9-dfb8-4767-aff8-1524bb04047e","pHszWG3wp179UVIj1kAzJ3",{"base":426},{"id":423,"rev":427},"8shuzYwFoOKusaQyp1Q6Le","post","2026-07-13T09:58:16Z",[431],"engineering",[433],{"_key":434,"_ref":435,"_type":92},"6655325c44f2","feeb4d3c-46b9-47cc-a997-7d8b487c7520",[437],{"_key":438,"_ref":439,"_type":92},"4fe50cf56b15","9613ea96-3dba-4c14-9ac9-2b5fe14eac53",[441,500,503,570,572,624,631,699,701,839,845,864,866,885,891,910,912,947,953,1024,1026,1069],{"_key":442,"_type":443,"text":444},"69d51ba0eb7f","textSection",[445,454,462,476,488],{"_key":446,"_type":130,"children":447,"markDefs":452,"style":453},"dc546114bad8",[448],{"_key":449,"_type":134,"marks":450,"text":451},"7ae30ff716a1",[],"Summary",[],"h2",{"_key":455,"_type":130,"children":456,"markDefs":461,"style":138},"35daf495c44a",[457],{"_key":458,"_type":134,"marks":459,"text":460},"fe2e6b919338",[],"On our journey toward training on a trillion tokens, we have learned what matters. In this post we share how we build:",[],{"_key":463,"_type":130,"children":464,"level":101,"listItem":474,"markDefs":475,"style":138},"92d0a7bcc16f",[465,470],{"_key":466,"_type":134,"marks":467,"text":469},"b439f69327b0",[468],"strong","A stable and efficient training framework.",{"_key":471,"_type":134,"marks":472,"text":473},"0b806bdb7063",[]," Enabling seamless training resumption after preemption or hardware failure, packing strategies to keep compute on real tokens, and end-to-end testing to catch regressions before they become expensive.","bullet",[],{"_key":477,"_type":130,"children":478,"level":101,"listItem":474,"markDefs":487,"style":138},"15b3772e2b21",[479,483],{"_key":480,"_type":134,"marks":481,"text":482},"f0908f44052a",[468],"Thorough data observability.",{"_key":484,"_type":134,"marks":485,"text":486},"09d749fd6437",[]," We monitor what actually reaches the model, not just the recipe: the realized mixture and per-source statistics, collected live during the run.",[],{"_key":489,"_type":130,"children":490,"level":101,"listItem":474,"markDefs":499,"style":138},"2045655f21e5",[491,495],{"_key":492,"_type":134,"marks":493,"text":494},"c0869ece124f",[468],"Trustworthy benchmarks.",{"_key":496,"_type":134,"marks":497,"text":498},"53af3e506dea",[]," We don't take a score at face value: aggressive decontamination of the training data against every benchmark, and recording the generations behind a number.",[],{"_key":501,"_type":502},"705ca7d0a22e","divider",{"_key":504,"_type":443,"text":505},"7c4a8e4216bf",[506,514,522,530,538,546,554,562],{"_key":507,"_type":130,"children":508,"markDefs":513,"style":453},"970c73cf1b5b",[509],{"_key":510,"_type":134,"marks":511,"text":512},"a8bc1dde2f73",[],"Introduction",[],{"_key":515,"_type":130,"children":516,"markDefs":521,"style":138},"ec6bcaa383e8",[517],{"_key":518,"_type":134,"marks":519,"text":520},"55a1bc359f8b",[],"Building frontier agentic reasoning models for clinical work means training a generalist model to navigate clinical workflows end-to-end: reasoning over long patient records, calling diagnostic tools, navigating radiology viewers, scrolling through whole-slide pathology images, and connecting what it learns from those tools back to a clinical question. These are tools that off-the-shelf foundation models never see during training and are not natively calibrated to operate.",[],{"_key":523,"_type":130,"children":524,"markDefs":529,"style":138},"e1bbd08160e0",[525],{"_key":526,"_type":134,"marks":527,"text":528},"148332c41fcc",[],"In a little over a year, we moved from training dense ~7B parameter multimodal vision-language models on a few hundred million tokens to 100B-token runs on 10-100x larger MoE models, with a roadmap toward trillion-token-scale continued pre-training (CPT). In our approach, the model architecture is treated as fixed. By benchmarking several open weight models that reported data sources and performances transparently, we settled on a strong base. This leaves the data as the primary variable we control in CPT and post-training. In what follows, we will therefore take a data-centric view on the stack that we’ve built.",[],{"_key":531,"_type":130,"children":532,"markDefs":537,"style":138},"0feb0297fe27",[533],{"_key":534,"_type":134,"marks":535,"text":536},"2f65f92a8251",[],"The first decision is the mixture itself: a three-way tug-of-war between installing new clinical knowledge, retaining the base's general skills, and warming up its latent ones. The rest is the engineering that delivers that mixture at scale:",[],{"_key":539,"_type":130,"children":540,"level":101,"listItem":474,"markDefs":545,"style":138},"899d88d4a02c",[541],{"_key":542,"_type":134,"marks":543,"text":544},"cf5869a8c799",[],"a robust training framework for 100B+ MoE models that tracks not just training progress but the dynamics, hardware utilization, and data statistics of a run",[],{"_key":547,"_type":130,"children":548,"level":101,"listItem":474,"markDefs":553,"style":138},"7ba6d98887b7",[549],{"_key":550,"_type":134,"marks":551,"text":552},"a0e19315943e",[],"efficient data streaming to focus the compute where it matters",[],{"_key":555,"_type":130,"children":556,"level":101,"listItem":474,"markDefs":561,"style":138},"3fa8ad03a873",[557],{"_key":558,"_type":134,"marks":559,"text":560},"2a27f23ca539",[],"observability to track down any failure and resume from a given state across data, model, and hardware (and to see what the model actually trained on)",[],{"_key":563,"_type":130,"children":564,"level":101,"listItem":474,"markDefs":569,"style":138},"4deb1bb0c509",[565],{"_key":566,"_type":134,"marks":567,"text":568},"3219294c411b",[],"an evaluation signal we can trust: broadly mapping the performance landscape we care about and rigorously decontaminating the test data",[],{"_key":571,"_type":502},"43567e705a3f",{"_key":573,"_type":443,"text":574},"c634445876a4",[575,583,600],{"_key":576,"_type":130,"children":577,"markDefs":582,"style":453},"106f6f5434aa",[578],{"_key":579,"_type":134,"marks":580,"text":581},"4faf2a1ff06e",[],"Mid-training as a three-way tug-of-war",[],{"_key":584,"_type":130,"children":585,"markDefs":599,"style":138},"85155a6b66aa",[586,590,595],{"_key":587,"_type":134,"marks":588,"text":589},"43cbeef9d998",[],"Rather than pre-training models from scratch, we start from capable models that have already been instruction-tuned and optimized for reasoning, and tool-use. We have previously verified that this yields stronger and better-performing models than starting from a checkpoint that was only pre-trained. However, a naïve continued pre-training on raw biomedical domain data would easily jeopardize the model’s learnt capabilities. To preserve these learnt skills, we craft a careful mix of raw, web-scale data (both text-only, image-captioning, and interleaved data), instruction-formatted Q&A and VQA, multi-turn conversations (with and without reasoning traces and images), as well as tool-use data. These instruction-formatted data components lean heavily on the general domain, while the raw, unformatted data is predominantly composed of domain-specific biomedical and clinical data intended to teach the model domain-specific knowledge. Technically, what we do is CPT, yet our data mixture moves this closer to what is often understood as mid-training, see e.g. in Liu et al. (2025)",{"_key":591,"_type":134,"marks":592,"text":594},"cb53f06731d0",[593],"sup","1",{"_key":596,"_type":134,"marks":597,"text":598},"4bbd257e7830",[],". Conventional CPT would rely exclusively on in-domain data and is often accompanied by a sharp performance drop in orthogonal domains – something that we aim to prevent by carefully crafting a balanced data mix.",[],{"_key":601,"_type":130,"children":602,"markDefs":620,"style":138},"3eab3a9d729e",[603,607,612,616],{"_key":604,"_type":134,"marks":605,"text":606},"e7924949fda1",[],"Thus, we view our CPT stage as a three-way tug-of-war between warmup, skill retention, and the target domain shift; see also ",{"_key":608,"_type":134,"marks":609,"text":611},"2d0705ed7503",[610],"5e7f12df0c22","PRISM",{"_key":613,"_type":134,"marks":614,"text":615},"0c8932053595",[593],"2",{"_key":617,"_type":134,"marks":618,"text":619},"e9e15b276cc4",[]," for an in-depth analysis of the mid-training paradigm. Here we will briefly outline each of these directions:",[621],{"_key":610,"_type":622,"href":623},"link","https://arxiv.org/abs/2603.17074",{"_key":625,"_type":626,"caption":627,"image":628},"3df31b978f06","imageSection","Mid-training as a tug-of-war between three data types, each defined by which two properties it combines. The circles are the properties (biomedical content, web-scale volume, instruction/tool-use/reasoning format); each data type sits in a pairwise overlap: domain shift (biomedical + web-scale) reinforces clinical knowledge, retention (web-scale + instruction) guards the base's general skills, warmup (biomedical + instruction) wakes latent clinical skills. Each pulls the mixture toward it with a force set by its abundance, so warmup, scarce and expensive to obtain, pulls with the thinnest rope. The triple overlap, representing web-scale biomedical instruction data, is not readily available and we reserve this for a future blog post on synthetic data generation for post-training.",{"_type":89,"asset":629},{"_ref":630,"_type":92},"image-c6b872846dd7f33141ab065174d0ecad477d3940-728x462-svg",{"_key":632,"_type":443,"text":633},"01cd7e1f72a0",[634,650,666,691],{"_key":635,"_type":130,"children":636,"markDefs":649,"style":138},"3fabde72f18d",[637,641,645],{"_key":638,"_type":134,"marks":639,"text":640},"13fd156cf704",[],"The base model already has latent clinical core skills (instruction following, medical reasoning, basic tool-use), but calibrated for consumer-facing chatbots rather than assisting in professional clinical workflows. ",{"_key":642,"_type":134,"marks":643,"text":644},"b305ca4180bd",[468],"Warmup",{"_key":646,"_type":134,"marks":647,"text":648},"e6936aee7d2e",[]," turns up the model's responsiveness on these skills by exposing it to clinical-shaped versions of them, e.g. reasoning traces injected into biomedical articles, or clinical VQA.",[],{"_key":651,"_type":130,"children":652,"markDefs":665,"style":138},"f0253901d30c",[653,657,661],{"_key":654,"_type":134,"marks":655,"text":656},"c22fb66d7184",[],"The second direction is ",{"_key":658,"_type":134,"marks":659,"text":660},"dff92df3f7ba",[468],"retention",{"_key":662,"_type":134,"marks":663,"text":664},"24e91c1880f0",[],": The base already knows English and a few other languages, math, code, and has some broad \"world knowledge\". We can't afford to erode any of that while chasing clinical capability. Catastrophic forgetting is the classic version of this risk. Our mixture is shaped accordingly: A large fraction of the data mix is general instruction-formatted data that uplifts the skills which otherwise would erode under the heavy weight of biomedical data.",[],{"_key":667,"_type":130,"children":668,"markDefs":690,"style":138},"d40f040fe896",[669,673,677,681,686],{"_key":670,"_type":134,"marks":671,"text":672},"65bc8e908242",[],"Finally, we adapt the model to the clinical domain by exposing it to the knowledge and structures of its target domain: biomedical terminology, the shape of EHRs, the literature, the conventions and guidelines of clinical reasoning. This is the ",{"_key":674,"_type":134,"marks":675,"text":676},"410e5284309c",[468],"domain shift",{"_key":678,"_type":134,"marks":679,"text":680},"26c2981fcf45",[]," we are targeting and what most people think of when they hear ",{"_key":682,"_type":134,"marks":683,"text":685},"747daafbab57",[684],"em","continued pre-training",{"_key":687,"_type":134,"marks":688,"text":689},"c17d0cac1ee9",[],": raw, web-scale domain data.",[],{"_key":692,"_type":130,"children":693,"markDefs":698,"style":138},"ba6a77785636",[694],{"_key":695,"_type":134,"marks":696,"text":697},"f85eef49fe42",[],"Each data type pulls the model in a different direction and striking the right balance to meet the criteria of a production-grade clinical model requires detailed understanding of the intended uses of the model, but equally important is a thorough understanding of the data mix in training. We will come back to this momentarily.",[],{"_key":700,"_type":502},"8360cdaf10a7",{"_key":702,"_type":443,"text":703},"3b0612e19a60",[704,712,720,728,736,748,760,768,776,784,800,812],{"_key":705,"_type":130,"children":706,"markDefs":711,"style":453},"2de0048034bc",[707],{"_key":708,"_type":134,"marks":709,"text":710},"872be5594ca6",[],"Building the foundation",[],{"_key":713,"_type":130,"children":714,"markDefs":719,"style":138},"e010aac9b769",[715],{"_key":716,"_type":134,"marks":717,"text":718},"cd7d3e113836",[],"Our aim for the training framework is that scaling up model size and data is a change of configuration, not a reinvention of the wheel. Everything above rests on it, so the infrastructure underneath has to be solid. Here’s what we concluded on.",[],{"_key":721,"_type":130,"children":722,"markDefs":727,"style":453},"0d5e1577609d",[723],{"_key":724,"_type":134,"marks":725,"text":726},"58f0fbcd59a2",[],"Training framework and data loading",[],{"_key":729,"_type":130,"children":730,"markDefs":735,"style":138},"2429fb33025d",[731],{"_key":732,"_type":134,"marks":733,"text":734},"8079b16ff675",[],"After evaluating the alternatives, we settled on building on top of NVIDIA's NeMo framework: Megatron-Bridge for training and Energon for data loading, respectively. What we needed was production-readiness at scale with enough integration surface to layer our own callbacks, logging, and recipe defaults on top, and both Megatron-Bridge and Energon fit: they are fully integrated from data loading, model training, and performance tracking, all the way to checkpoint resumption, and at the same time allow us to build on top, add custom training logic and logging where we need it. As a bonus, NVIDIA's NeMo comes with a range of low-level hardware accelerations that result in 5x faster training iterations compared to our old research training framework built on top of PyTorch Lightning.",[],{"_key":737,"_type":130,"children":738,"markDefs":747,"style":138},"338523a84915",[739,743],{"_key":740,"_type":134,"marks":741,"text":742},"f0a387393db2",[468],"Tests at every scale of integration.",{"_key":744,"_type":134,"marks":745,"text":746},"3e7c8c20e52c",[]," While unit tests ensure that every component in the pipeline behaves as expected, unforeseen long-range interactions both within one team and across team boundaries can lead to undesirable outcomes. Therefore, we test end-to-end: For any change to the data or the code we trigger integration tests for the entire pipeline including short training runs and full iterations of the affected datasets to ensure continuity in our production training runs.",[],{"_key":749,"_type":130,"children":750,"markDefs":759,"style":138},"9b4f97296995",[751,755],{"_key":752,"_type":134,"marks":753,"text":754},"9e0baf1e6048",[468],"Byte-identical resumability.",{"_key":756,"_type":134,"marks":757,"text":758},"f62d3f6b2bfe",[]," A multi-week run will be interrupted at some point, either by preemption, hardware failure, or a planned restart. This is where the NeMo universe shines: We persist the full dataloader state alongside every model checkpoint, thereby ensuring byte-identical behavior between the uninterrupted run and its continuation.",[],{"_key":761,"_type":130,"children":762,"markDefs":767,"style":453},"2bf9ba0cb06b",[763],{"_key":764,"_type":134,"marks":765,"text":766},"d7a762e1c697",[],"Packing and token efficiency",[],{"_key":769,"_type":130,"children":770,"markDefs":775,"style":138},"abbfbc797779",[771],{"_key":772,"_type":134,"marks":773,"text":774},"63d23b52b4e4",[],"Why production-readiness matters at this scale is best illustrated by putting numbers to it. One example that's easy to overlook is sequence packing.",[],{"_key":777,"_type":130,"children":778,"markDefs":783,"style":138},"29813ce26ead",[779],{"_key":780,"_type":134,"marks":781,"text":782},"e6d75aae884f",[],"Biomedical and clinical corpora are wildly heterogeneous in document length: a PubMed abstract runs around 200 tokens, clinical notes vary broadly (500-3K), full papers and tool traces easily exceed 10K. Packing sequentially loaded samples from such a heterogeneous distribution into a standard pre-training sequence of length, say, 4k tokens, inevitably results in a substantial amount of padding tokens: long sequences are truncated, long and short sequences get combined into sub-optimal combinations. To alleviate this, we use a greedy Knapsack algorithm that selects packing candidates optimally from a pre-loaded buffer. Moreover, we choose a packing strategy that minimizes the padding content. Three packing regimes are worth comparing:",[],{"_key":785,"_type":130,"children":786,"level":101,"listItem":474,"markDefs":799,"style":138},"374fc10da694",[787,791,795],{"_key":788,"_type":134,"marks":789,"text":790},"5f8528990b14",[468],"No packing.",{"_key":792,"_type":134,"marks":793,"text":794},"6c7307634e50",[]," One document per sequence, padded out to the context length. On a biomedical document-length mix at 4K context, realistic padding waste is ",{"_key":796,"_type":134,"marks":797,"text":798},"00595e8629fc",[468],"30-40%.",[],{"_key":801,"_type":130,"children":802,"level":101,"listItem":474,"markDefs":811,"style":138},"575ade0b68a1",[803,807],{"_key":804,"_type":134,"marks":805,"text":806},"3084a27232c9",[468],"Naïve packing.",{"_key":808,"_type":134,"marks":809,"text":810},"aacf4e83e498",[]," Concatenate samples into one sequence, no attention masking. Padding waste drops to near zero, but attention across document boundaries can contaminate the gradient signal and has sub-optimal memory scaling.",[],{"_key":813,"_type":130,"children":814,"level":101,"listItem":474,"markDefs":836,"style":138},"5a74c027d94a",[815,819,823,828,832],{"_key":816,"_type":134,"marks":817,"text":818},"7cc761332000",[468],"Boundary-aware packing.",{"_key":820,"_type":134,"marks":821,"text":822},"74f97f6436e5",[]," Concatenate samples into a single sequence, but mask attention so each token only attends to others within its own document. This is the approach chosen by us and other frontier labs, see e.g. ",{"_key":824,"_type":134,"marks":825,"text":827},"cd23054625a3",[826],"1d1c311fb0db","DeepSeek-V4",{"_key":829,"_type":134,"marks":830,"text":831},"21210b1e6afa",[593],"3",{"_key":833,"_type":134,"marks":834,"text":835},"89641ab631f8",[],". It guarantees near-zero padding waste and each document is treated as an individual sequence.",[837],{"_key":826,"_type":622,"href":838},"https://arxiv.org/abs/2606.19348",{"_key":840,"_type":626,"caption":841,"image":842},"445ea7d6238d","Top: six documents of different lengths. One document per sequence (left) leaves 30 - 40% of tokens as padding (note that for multi-node training, a global sequence length must be fixed); concatenating the documents (right) removes most of it, with a small residual if sequences are split only at paragraph or sentence boundaries. Bottom: the causal attention matrix over one packed sequence of three documents (A, B, C). Within-document attention (blue) is identical in both cases; naïve packing (left) additionally attends across document boundaries (gradient contamination and suboptimal memory scaling), while boundary-aware packing (right) masks those cross-document positions, so each document is treated as an individual sequence.",{"_type":89,"asset":843},{"_ref":844,"_type":92},"image-fd14a1c0d14abf45a5e9c08332953644f2cf7535-960x780-svg",{"_key":846,"_type":443,"text":847},"40929a7023b1",[848],{"_key":849,"_type":130,"children":850,"markDefs":863,"style":138},"d19cc193a915",[851,855,859],{"_key":852,"_type":134,"marks":853,"text":854},"20935da01ddb",[],"In GPU-hours this would mean the following. The public Qwen3.5-VL 122B-A10B SFT recipe runs at global batch size 36, with a maximum sequence length of 4,096, for 300k steps: roughly 44B training tokens, or about 2'100 GPU-hours on 48 H100s at 35% MFU. At a worst-case 40% padding waste, ~840 of those GPU-hours are spent processing padded positions that contribute no useful training signal. Scaling the same math up, a 100B-token run wastes around 1'900 GPU-hours per recipe ablation, and the 1T-token target wastes around 19'000. That's roughly ",{"_key":856,"_type":134,"marks":857,"text":858},"92e681c9e784",[468],"$76K per run",{"_key":860,"_type":134,"marks":861,"text":862},"ed6e374b63f6",[]," at typical H100 cloud pricing of around 4$/GPU-hour.",[],{"_key":865,"_type":502},"9a8765b71e4c",{"_key":867,"_type":443,"text":868},"8d3ef62cd2fb",[869,877],{"_key":870,"_type":130,"children":871,"markDefs":876,"style":453},"2f1d3ec018ee",[872],{"_key":873,"_type":134,"marks":874,"text":875},"449673190e49",[],"What did the model actually see?",[],{"_key":878,"_type":130,"children":879,"markDefs":884,"style":138},"c935a02a3b9e",[880],{"_key":881,"_type":134,"marks":882,"text":883},"2aa229c788ff",[],"One subtlety matters here specifically since the realized and planned data mixtures can deviate substantially: The planned mix is a list of sources with their target shares, a set of filter parameters, and a training schedule. However, what reaches the model might deviate substantially from this planned mix. We illustrate this with the figure below (left panel). A declared (sample level) recipe would contain 71% web-scale biomedical text data; however, after transformations and filters are applied, sequences have been packed, and bad or uninformative samples removed (think of corrupted images that previously weren’t caught), the true token level mixture looks quite different, with a sharply decreased share of the web scale text dataset. Further looking at the loss token share per dataset, the tension relaxes but still strongly differs from the declared mix.",[],{"_key":886,"_type":626,"caption":887,"image":888},"e7e060cc1dba","A real run (~20B tokens; sources anonymized), counted three ways. Left: the same realized data weighted by sample, by token, and by loss-token. By sample it matches the declared blend (71% biomedical text / 10% biomedical image-text / 19% skill additions); by token, biomedical text falls to 43% as token-dense skill data more than doubles its share; by loss-token (prompt, system, and image tokens masked out) biomedical text climbs back to 51%, while masked tool-use data collapses from 12% to 5%, far less than its token share suggests. Right: sources differ in size, so a small tool-use subset is cycled ~9 times before the largest corpus finishes a fifth of one pass.",{"_type":89,"asset":889},{"_ref":890,"_type":92},"image-22de86cabced38eb9772f42314b21a3c93b7acdb-1120x460-svg",{"_key":892,"_type":443,"text":893},"be90546e868a",[894,902],{"_key":895,"_type":130,"children":896,"markDefs":901,"style":138},"89efc4513b11",[897],{"_key":898,"_type":134,"marks":899,"text":900},"9cf280f773d2",[],"Another aspect that is crucial to monitor is the number of times each dataset in the mix is seen. A first best judgement blend will be altered through the aforementioned dynamics and consequently we might be seeing the same samples more than intended. The right panel of the figure illustrates this, where a small tool-use dataset is heavily oversampled. The mechanism is as follows: While the dataset might have been correctly weighted by token share (tool-use data tends to come with fewer samples but very long sequences), many of the samples would exceed the context window of the initial stage of CPT already with the system message and user prompt, which depending on the training configuration may leave no tokens relevant for gradient calculation at all and thus are filtered on the fly. The resulting over-representation of the dataset in the mix requires a subsequent re-adjustment.",[],{"_key":903,"_type":130,"children":904,"markDefs":909,"style":138},"4b03cf391aaf",[905],{"_key":906,"_type":134,"marks":907,"text":908},"335716809993",[],"The cost of being wrong grows with run size. At the trillion-token scale we're working toward, a 1% drift in the realized mixture is 10B tokens of unintended training spent on the wrong distribution (about the entire budget of a respectable open research run). Our token-level observability stack ensures that any drift is visible at training time and not hidden in a silent performance drift in evaluation.",[],{"_key":911,"_type":502},"c70574fa1114",{"_key":913,"_type":443,"text":914},"78af5b6faf74",[915,923,931,939],{"_key":916,"_type":130,"children":917,"markDefs":922,"style":453},"7aae45ba6e15",[918],{"_key":919,"_type":134,"marks":920,"text":921},"85ca49ba3bf3",[],"A trustworthy evaluation signal",[],{"_key":924,"_type":130,"children":925,"markDefs":930,"style":138},"4be2a7958365",[926],{"_key":927,"_type":134,"marks":928,"text":929},"dd6e8aa47276",[],"Checkpoint validation is our main signal for whether the upstream work paid off, so it carries the same engineering investment as training and data.",[],{"_key":932,"_type":130,"children":933,"markDefs":938,"style":138},"22e53ea0f1e8",[934],{"_key":935,"_type":134,"marks":936,"text":937},"f0c25d2f4d9c",[],"Contamination is a sharper risk for us than for a research model. A research model is judged on a fixed, public benchmark set it can decontaminate against; our models are judged by users on clinical tasks we may not have anticipated. Therefore, we evaluate against both critical public benchmarks and internal private ones, and keep adding and revising them as the product surface grows. One uncaught test-set leak turns that signal into a misleading one.",[],{"_key":940,"_type":130,"children":941,"markDefs":946,"style":138},"d4e39d0546a2",[942],{"_key":943,"_type":134,"marks":944,"text":945},"77274cdcdeb2",[],"Therefore, we deduplicate every training source against every active benchmark. Exact matching runs on every pair; we add the costly fuzzy and semantic passes where source and benchmark draw on shared material. Overlap on the rest is tracked as a tripwire, escalating a pair if it climbs above a threshold.",[],{"_key":948,"_type":626,"caption":949,"image":950},"a508836edf99","Leakage risk varies across (training source × benchmark) pairs, so we target the expensive matching rather than spread it evenly. Every pair is deduplicated by exact match; where a source and benchmark draw on shared material (high-risk pairs) we add the full fuzzy and semantic stack (dense-embedding matching for images). Overlap on the rest is tracked as a tripwire and escalated if it climbs. New sources and benchmarks are matched against everything active as they come online.",{"_type":89,"asset":951},{"_ref":952,"_type":92},"image-7ef14539e1371164e7cf6dbda271faf29106a01e-780x470-svg",{"_key":954,"_type":443,"text":955},"e04dea550fa1",[956,964,976,996,1016],{"_key":957,"_type":130,"children":958,"markDefs":963,"style":138},"a74f34b0467f",[959],{"_key":960,"_type":134,"marks":961,"text":962},"a6f511a6172f",[],"A trustworthy signal needs two things: a benchmark the model has not already seen, and a way to read why a score moved. For the first, we escalate the matcher, each step catching what the one before it cannot:",[],{"_key":965,"_type":130,"children":966,"level":101,"listItem":474,"markDefs":975,"style":138},"8355dc6045b8",[967,971],{"_key":968,"_type":134,"marks":969,"text":970},"883ddbe07a8c",[468],"Exact match.",{"_key":972,"_type":134,"marks":973,"text":974},"b97e43c51e87",[]," Verbatim copies fall out of a hash match, run on every (source × benchmark) pair.",[],{"_key":977,"_type":130,"children":978,"level":101,"listItem":474,"markDefs":995,"style":138},"0dba7b474d3d",[979,983,987,991],{"_key":980,"_type":134,"marks":981,"text":982},"5296869c12eb",[468],"Fuzzy matching.",{"_key":984,"_type":134,"marks":985,"text":986},"4789fc6e165f",[]," Reformatted and reworded copies survive it: MinHash-LSH surfaces near-duplicates by Jaccard overlap",{"_key":988,"_type":134,"marks":989,"text":990},"b025fbb6c976",[593],"4",{"_key":992,"_type":134,"marks":993,"text":994},"5228c870ab07",[],".",[],{"_key":997,"_type":130,"children":998,"level":101,"listItem":474,"markDefs":1015,"style":138},"e717ab4975c6",[999,1003,1007,1011],{"_key":1000,"_type":134,"marks":1001,"text":1002},"1c449255b409",[468],"Semantic matching.",{"_key":1004,"_type":134,"marks":1005,"text":1006},"407039cff2af",[]," What survives fuzzy is caught by embeddings: a cosine pass over text",{"_key":1008,"_type":134,"marks":1009,"text":1010},"9f4a886e94d7",[593],"5",{"_key":1012,"_type":134,"marks":1013,"text":1014},"f0f305801ae8",[],", and dense image embeddings with approximate nearest-neighbour search that flag a training image answering a benchmark question even when its pixels differ. Both carry a higher false-positive cost, so we reserve them for pairs where a source and benchmark share material, as illustrated in the figure.",[],{"_key":1017,"_type":130,"children":1018,"markDefs":1023,"style":138},"a130dec621f4",[1019],{"_key":1020,"_type":134,"marks":1021,"text":1022},"0c39c1608f5c",[],"For the second point, we store all generations and not just the score. We pull actual generations from each checkpoint and inspect a few of them by hand and with an LLM judge at scale, because a drop in a benchmark has many possible causes a single number hides. A common example is a drop in performance, where the model gives the right answer in the wrong format, so the failure is instruction following, not knowledge. A score alone cannot tell the two apart.",[],{"_key":1025,"_type":502},"0fa9e390015a",{"_key":1027,"_type":443,"text":1028},"204aae08f9bd",[1029,1037,1045,1061],{"_key":1030,"_type":130,"children":1031,"markDefs":1036,"style":453},"dd7d700a4605",[1032],{"_key":1033,"_type":134,"marks":1034,"text":1035},"3111078cd748",[],"Conclusion and outlook",[],{"_key":1038,"_type":130,"children":1039,"markDefs":1044,"style":138},"d6a2d91f522f",[1040],{"_key":1041,"_type":134,"marks":1042,"text":1043},"4dd4e4410448",[],"Every stage in this post acts on the same data mixture: the framework and packing deliver it, observability records what the model actually trained on, and decontaminated evaluation shows whether it worked. Each component is built to hold as runs scale, so moving from today's 100B-token runs to the trillion-token target on 100B+ MoE models means changing the configuration, not rewriting the stack. What remains is to compose them into a single loop.",[],{"_key":1046,"_type":130,"children":1047,"markDefs":1060,"style":138},"8ac1185c23f4",[1048,1052,1056],{"_key":1049,"_type":134,"marks":1050,"text":1051},"1de81c4b8bf6",[],"That composition is what we are building toward in kaiko's ",{"_key":1053,"_type":134,"marks":1054,"text":1055},"60d25eb0e4a7",[468],"AI Factory",{"_key":1057,"_type":134,"marks":1058,"text":1059},"6552009fc270",[],", a pipeline where evaluation results feed back into the recipe directly, the recipe as a versioned, structured artifact and the observability primitives from the preceding sections making each stage's output legible to the next. The orchestration already runs on Dagster, workloads are distributed via Ray, and Kubernetes provisions and schedules the resources. The components exist today (framework, loader, packing, decontamination, a trustworthy evaluation signal, an emerging recipe artifact). Composing them into a single closed loop is the work ahead, and the subject of a future deep dive.",[],{"_key":1062,"_type":130,"children":1063,"markDefs":1068,"style":138},"f6c7f0283de9",[1064],{"_key":1065,"_type":134,"marks":1066,"text":1067},"1a31159132d3",[],"None of these practices are novel on their own. The literature has assembled most of the pieces. What rarely makes it into a paper are the steps behind a headline: preempted and crashed runs, the mixture drifting from the recipe, or a dataset silently cycled nine times. With this article, we want to shed some light on what it takes to move toward the trillion-token frontier.",[],{"_key":1070,"_type":1071,"references":1072},"81a6af91cf49","referencesSection",[1073,1074,1075,1076,1077],"Liu et al. 2025, Midtraining Bridges Pretraining and Posttraining Distributions, \nhttps://arxiv.org/abs/2510.14865","PRISM: Demystifying Retention and Interaction in Mid-Training, https://arxiv.org/abs/2603.17074","DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence\nhttps://arxiv.org/abs/2606.19348","Lee et al. 2021 Deduplicating Training Data Makes Language Models Better, https://arxiv.org/abs/2107.06499 ","Abbas et al. 2023, SemDeDup: Data-efficient learning at web-scale through semantic deduplication, https://arxiv.org/abs/2303.09540","2026-07-09T16:12:54.910Z",{"_type":89,"asset":1080},{"_ref":1081,"_type":92},"image-d9a09026393fbffde93f88b44b36fe2f68e105d2-2010x938-png",[1083],{"_key":1084,"_ref":1085,"_type":92},"27a23ce5e680","38a85cd4-0666-4e99-b794-b9b8371c12e9",{"_type":1087,"current":1088},"slug","Training-clinical-reasoning-model-at-scale","Toward training at the trillion-token scale: Why reliability, stability, and observability matter and how we engineer them.",[1091,1094,1097,1100],{"_key":1092,"_ref":1093,"_type":92},"3b76795f501c","ffc879d7-9be3-47d1-bda5-5817c9d7cf8c",{"_key":1095,"_ref":1096,"_type":92},"c2dc7fdd784c","0e7a76c9-607b-4b29-b0f5-48940b955680",{"_key":1098,"_ref":1099,"_type":92},"a476af4cdeb6","038e94a3-7be4-458d-972c-d3b572c832a0",{"_key":1101,"_ref":1102,"_type":92},"279baccd7561","5f860d5a-1a0d-4041-ab25-eb07132ec4bd",{"_createdAt":1104,"_id":1105,"_rev":1106,"_system":1107,"_type":428,"_updatedAt":1110,"area":1111,"authors":1113,"categories":1120,"content":1124,"date":1539,"featuredImage":1540,"isHidden":94,"relatedPosts":1543,"slug":1550,"subHeading":1552,"title":1553,"topics":1554},"2026-06-23T16:59:31Z","b2ccb769-6e76-4b7a-8ebb-a93af0ea508c","KOW9SnXoBiNxp6TLs1rRaS",{"base":1108},{"id":1105,"rev":1109},"V49ZwwEaMrto2Vw8K0DYWe","2026-06-25T14:05:17Z",[1112],"medical",[1114,1117],{"_key":1115,"_ref":1116,"_type":92},"2740c82aa1af","9864f3e2-7a57-471f-9f93-870c1881908c",{"_key":1118,"_ref":1119,"_type":92},"5b87736bc4f7","7adade4c-f729-4230-aacf-dd4e9a61824f",[1121],{"_key":1122,"_ref":1123,"_type":92},"5eb8111180e9","f3f4fc6c-726a-4c2c-97b5-8e6859bb4741",[1125,1152,1157,1159,1202,1204,1236,1240,1271,1277,1304,1307,1326,1329,1347,1350,1389,1392,1410,1416,1418,1457,1459,1519,1527],{"_key":1126,"_type":443,"text":1127},"1cdc03be6cb2",[1128,1136],{"_key":1129,"_type":130,"children":1130,"markDefs":1135,"style":138},"27fbe86fcc5c",[1131],{"_key":1132,"_type":134,"marks":1133,"text":1134},"02403460d103",[],"Artificial intelligence is finding its way into hospitals, but not evenly. Some departments are further along than others. Pathology is one where the potential is already visible and, in some hospitals, already at work. The reasons are not hard to find: a growing volume of diagnoses, a shrinking pool of specialists, and a type of work that is well-suited to what AI does best.",[],{"_key":1137,"_type":130,"children":1138,"markDefs":1151,"style":138},"f4a8793a8545",[1139,1143,1147],{"_key":1140,"_type":134,"marks":1141,"text":1142},"b7a48740df6a",[],"Hugo Horlings is one of the pathologists working at that intersection every day. Based at the ",{"_key":1144,"_type":134,"marks":1145,"text":1146},"4d26332f0dd1",[684],"Netherlands Cancer Institute",{"_key":1148,"_type":134,"marks":1149,"text":1150},"4ab2985c55aa",[]," in Amsterdam, one of Europe's leading cancer centres, he spends his days analysing tissue samples, identifying cancer cells, and helping determine how patients should be treated.",[],{"_key":1153,"_type":626,"image":1154},"a428538f9f1f",{"_type":89,"asset":1155},{"_ref":1156,"_type":92},"image-e7d43b8b68f29ca1d944ef81d4ab326c39fe4ffc-1312x662-png",{"_key":1158,"_type":502},"932992bc7542",{"_key":1160,"_type":443,"text":1161},"986434a75852",[1162,1170,1178,1186,1194],{"_key":1163,"_type":130,"children":1164,"markDefs":1169,"style":453},"f6617f76c084",[1165],{"_key":1166,"_type":134,"marks":1167,"text":1168},"a4ad3cbe9daa",[468],"AI in the clinical setting: why it matters",[],{"_key":1171,"_type":130,"children":1172,"markDefs":1177,"style":138},"7876147ac5c9",[1173],{"_key":1174,"_type":134,"marks":1175,"text":1176},"5b5724e4a7d0",[],"Pathology is a clear-cut example of a highly precise and demanding job that requires expert judgement, yet there are fewer experts to do it.",[],{"_key":1179,"_type":130,"children":1180,"markDefs":1185,"style":138},"247dc38540df",[1181],{"_key":1182,"_type":134,"marks":1183,"text":1184},"1f1eaabfb147",[],"A single biopsy contains roughly a million cells; pathologists can’t count them individually, nobody does, according to Hugo. 'First you do a first-eye sweep,' he says, 'and based on that you make an estimation.' This requires solid expert judgment, built over years. And the volume of work that demands it is growing.",[],{"_key":1187,"_type":130,"children":1188,"markDefs":1193,"style":138},"21fd02b32695",[1189],{"_key":1190,"_type":134,"marks":1191,"text":1192},"5deaf5dab36c",[],"To further sketch the situation, Hugo continues: “On top of that, the number of cancer diagnoses is rising, but the number of pathologists is not. At the time of this interview, in spring 2026, there were 22 open vacancies for pathologists in the Netherlands, a figure Hugo describes as the highest it has ever been.",[],{"_key":1195,"_type":130,"children":1196,"markDefs":1201,"style":138},"4ac01470ede0",[1197],{"_key":1198,"_type":134,"marks":1199,"text":1200},"edea7099be0b",[],"\"We have to do more work with the same or even fewer pathologists,\" Hugo says. And the same pressure is building across nursing and clinical support roles.",[],{"_key":1203,"_type":502},"e8e121eef4f5",{"_key":1205,"_type":443,"text":1206},"a7d114060d46",[1207,1215,1222,1229],{"_key":1208,"_type":130,"children":1209,"markDefs":1214,"style":453},"61341a75ce73",[1210],{"_key":1211,"_type":134,"marks":1212,"text":1213},"717b116f1332",[],"What AI actually does",[],{"_key":1216,"_type":130,"children":1217,"markDefs":1221,"style":138},"046075a3b890",[1218],{"_key":1211,"_type":134,"marks":1219,"text":1220},[],"In practice, we know that AI helps clinicians make better use of the time they have. In Hugo’s case: take a stack of 15 pathology slides. Today, a pathologist reviews each one. With AI pre-screening those slides and flagging the ones that require closer attention, that same pathologist might only need to closely examine two or three.",[],{"_key":1223,"_type":130,"children":1224,"markDefs":1228,"style":138},"bb409430d5be",[1225],{"_key":1211,"_type":134,"marks":1226,"text":1227},[],"They still see every slide. They still make every call. But their focus goes where it’s most needed.",[],{"_key":1230,"_type":130,"children":1231,"markDefs":1235,"style":138},"2858a7f7c9f8",[1232],{"_key":1211,"_type":134,"marks":1233,"text":1234},[],"But beyond speed, AI is beginning to do something noteworthy. It sees things differently. A pathologist looking at a tissue sample immediately takes in a million cells, making pattern-based judgments built on years of training. AI can count every single one of those cells and then ask questions a human eye simply can’t answer at scale. Are these immune cells clustered near the tumour? Are they touching it? Where exactly are they located? That kind of spatial detail, Hugo explains, is opening new territory entirely: ",[],{"_key":1237,"_type":1238,"quote":1239},"6f07dd88aae6","pullquoteSection","An AI model can count those immune cells reliably and quickly. More than that, it can map where they sit.",{"_key":1241,"_type":443,"text":1242},"1d3496c376cc",[1243],{"_key":1244,"_type":130,"children":1245,"markDefs":1270,"style":138},"af3129381841",[1246,1250,1253,1257,1260,1264,1267],{"_key":1247,"_type":134,"marks":1248,"text":1249},"f9d54a4977ff",[],"\nThis is particularly important for Hugo and his team. They study immune cells that gather near tumours as part of the body's natural defences, known as tumour-infiltrating lymphocytes",{"_key":1251,"_type":134,"marks":1252,"text":594},"759fd7a3f267",[593],{"_key":1254,"_type":134,"marks":1255,"text":1256},"5b6726e947e6",[],". In earlier research",{"_key":1258,"_type":134,"marks":1259,"text":615},"8cda06b45a30",[593],{"_key":1261,"_type":134,"marks":1262,"text":1263},"935be9f57669",[],", he and his colleague, Marleen Kok and team, focused on a specific type of breast cancer: small triple-negative tumours, which are typically treated with chemotherapy. What they found was unexpected. Patients with a high density of these immune cells had survival rates equivalent to those of patients without cancer, even 15 years after diagnosis",{"_key":1265,"_type":134,"marks":1266,"text":615},"c519d54a3986",[593],{"_key":1268,"_type":134,"marks":1269,"text":994},"ca90ac586713",[],[],{"_key":1272,"_type":626,"caption":1273,"image":1274},"3cefa2ad45eb","Example of a tissue sample showing tumour-infiltrating lymphocytes (TILs) under the microscope. The small, dark cells are the immune cells Hugo's team studies. Example from WebPathology [4]",{"_type":89,"asset":1275},{"_ref":1276,"_type":92},"image-824837c439cca02a9dcbd768a147b744fb4bad2b-1288x774-png",{"_key":1278,"_type":443,"text":1279},"50100fd2d310",[1280,1296],{"_key":1281,"_type":130,"children":1282,"markDefs":1295,"style":138},"019297fc8375",[1283,1287,1291],{"_key":1284,"_type":134,"marks":1285,"text":1286},"3e6c8332cf53",[],"The implication is significant. For some of these patients, chemotherapy may not be necessary at all, sparing them treatments with serious side effects they do not need. Whether that is indeed the case is now being investigated in the OPTImaL Trial",{"_key":1288,"_type":134,"marks":1289,"text":1290},"8182ade5e437",[593],"1,3",{"_key":1292,"_type":134,"marks":1293,"text":1294},"902a8fbe6f9a",[],", supported by the Netherlands Cancer Institute.",[],{"_key":1297,"_type":130,"children":1298,"markDefs":1303,"style":138},"d6eafd84687f",[1299],{"_key":1300,"_type":134,"marks":1301,"text":1302},"73eebe373e1c",[],"While this clinical insight exists without AI, the challenge is making it usable at scale. Across the Netherlands, 450 pathologists were needed to consistently identify and count these immune cells. Manually, that is slow and variable. With AI, it becomes tractable, according to Hugo:",[],{"_key":1305,"_type":1238,"quote":1306},"a5cc393f9a07","You can now ask the computer: where are these cells localized? How many of them are touching a tumour cell? Spatial information is reaching the next level, something we cannot see or count with our human eyes.",{"_key":1308,"_type":443,"text":1309},"350a7b081567",[1310,1318],{"_key":1311,"_type":130,"children":1312,"markDefs":1317,"style":453},"1e76ad96ea7a",[1313],{"_key":1314,"_type":134,"marks":1315,"text":1316},"e7200d925c19",[468],"AI and what it means for the role of the clinician",[],{"_key":1319,"_type":130,"children":1320,"markDefs":1325,"style":138},"0dfa0ac387b3",[1321],{"_key":1322,"_type":134,"marks":1323,"text":1324},"ef7fd6de2923",[],"There’s a version of the AI conversation in healthcare that sounds like science fiction: autonomous systems making diagnoses, algorithms replacing specialists, machines doing the work of doctors. Hugo is quick to set that aside.",[],{"_key":1327,"_type":1238,"quote":1328},"bfb9f03f29f9","I really see AI as an assistant.",{"_key":1330,"_type":443,"text":1331},"0647a343c4d2",[1332,1340],{"_key":1333,"_type":130,"children":1334,"markDefs":1339,"style":138},"ed694ca6f631",[1335],{"_key":1336,"_type":134,"marks":1337,"text":1338},"5a52a4eb71ef",[],"Under the EU AI Act, fully autonomous AI decision-making in clinical settings isn’t permitted, and Hugo agrees that this is appropriate. These models are still developing, and we’re still learning how to work with them. The human stays in the loop, and Hugo emphasizes that it should not only be the case because it is a regulatory formality, but more so because good medical practice demands it. ",[],{"_key":1341,"_type":130,"children":1342,"markDefs":1346,"style":138},"c73773f968cb",[1343],{"_key":1336,"_type":134,"marks":1344,"text":1345},[],"The patient who changes their mind. The family situation that shifts the calculation. The preference that exists nowhere in any record. No system captures that.",[],{"_key":1348,"_type":1238,"quote":1349},"2f8433e70c39","There are far more nuances than a computer can foresee. The decision always stays with the human; keeping the human in the loop is how responsible AI in medicine works",{"_key":1351,"_type":443,"text":1352},"54753c5104b6",[1353,1361,1368,1375,1382],{"_key":1354,"_type":130,"children":1355,"markDefs":1360,"style":453},"07545462cf34",[1356],{"_key":1357,"_type":134,"marks":1358,"text":1359},"5d3fe3c8abac",[],"‘From will it replace me’ to ‘what will it save me’",[],{"_key":1362,"_type":130,"children":1363,"markDefs":1367,"style":138},"01959ed6473e",[1364],{"_key":1357,"_type":134,"marks":1365,"text":1366},[],"A few years ago, the dominant fear Hugo heard from colleagues was straightforward: will AI replace me? He does not hear that much anymore, the conversation has moved.",[],{"_key":1369,"_type":130,"children":1370,"markDefs":1374,"style":138},"1ab4d9504bdb",[1371],{"_key":1357,"_type":134,"marks":1372,"text":1373},[684],"\"Now it is more: what will it actually save me? And the answer is quite a lot of time.\"",[],{"_key":1376,"_type":130,"children":1377,"markDefs":1381,"style":138},"6728a4d40b91",[1378],{"_key":1357,"_type":134,"marks":1379,"text":1380},[],"For many clinicians who have started experimenting, the answer turns out to be significant. Quite a lot of time, it turns out.",[],{"_key":1383,"_type":130,"children":1384,"markDefs":1388,"style":138},"0cbf22d00812",[1385],{"_key":1357,"_type":134,"marks":1386,"text":1387},[],"\nAccording to Hugo, there is a subtler change worth paying attention to: \n",[],{"_key":1390,"_type":1238,"quote":1391},"91b77a470894","AI does not just take tasks off a clinician's plate. It sometimes nudges you to look again. When a system flags something unexpected or surfaces a possibility, you had not considered, you check it and think again.",{"_key":1393,"_type":443,"text":1394},"1658573e1c66",[1395,1403],{"_key":1396,"_type":130,"children":1397,"markDefs":1402,"style":138},"b57526186587",[1398],{"_key":1399,"_type":134,"marks":1400,"text":1401},"3ae3ccc13dbc",[],"Hugo compares this to the introduction of structured checklists in aviation. Plane crashes dropped significantly when pilots started using them. Surgeons adopted the same approach in operating rooms, confirming before and after every procedure that nothing had been missed, and complications decreased. The checklist did not replace expertise; it caught what expertise, under pressure, sometimes misses.",[],{"_key":1404,"_type":130,"children":1405,"markDefs":1409,"style":138},"dfd0e2d92c7b",[1406],{"_key":1399,"_type":134,"marks":1407,"text":1408},[],"Hugo thinks AI could work similarly for diagnosis. A prompt to check again. A flag on something you might have moved past. Less room for one pathologist to give an estimation of five percent and another to say ten.",[],{"_key":1411,"_type":626,"caption":1412,"image":1413},"2da2c7909b75","Example of an aviation checklist, from the Flight Safety Foundation [5]",{"_type":89,"asset":1414},{"_ref":1415,"_type":92},"image-cdceb3a111ff8d1694c9e3895d6c25c845287f4e-824x550-jpg",{"_key":1417,"_type":502},"f935ca104639",{"_key":1419,"_type":443,"text":1420},"697fcb2cc769",[1421,1429,1436,1443,1450],{"_key":1422,"_type":130,"children":1423,"markDefs":1428,"style":453},"792b97e01e65",[1424],{"_key":1425,"_type":134,"marks":1426,"text":1427},"a6ed0ee36c93",[],"Sharing is multiplying",[],{"_key":1430,"_type":130,"children":1431,"markDefs":1435,"style":138},"deb4c1763a45",[1432],{"_key":1425,"_type":134,"marks":1433,"text":1434},[],"Near the end of the conversation, Hugo reaches for something his father always says. When you share something, you multiply. And this also accounts for knowledge.",[],{"_key":1437,"_type":130,"children":1438,"markDefs":1442,"style":138},"e1f225eca215",[1439],{"_key":1425,"_type":134,"marks":1440,"text":1441},[],"He applies it to what he sees as the only viable path forward. Hospitals cannot do this alone. They do not have the technical capacity, the legal infrastructure, or the bandwidth. Technology companies cannot do it alone either. They do not have the domain knowledge, the clinical understanding, or the trust that comes from years inside a department.",[],{"_key":1444,"_type":130,"children":1445,"markDefs":1449,"style":138},"47750c0924e3",[1446],{"_key":1425,"_type":134,"marks":1447,"text":1448},[],"“Bring that together,” he says, “and you get something truly valuable.” It is a simple idea. But in a field where hospitals and technology companies have spent years talking past each other, it is also the hardest thing to get right. For those ready to take a first step, his advice is characteristically direct: start small, expect the decisions to multiply rather than disappear, and do not try to build it alone.",[],{"_key":1451,"_type":130,"children":1452,"markDefs":1456,"style":138},"20aaa57e7508",[1453],{"_key":1425,"_type":134,"marks":1454,"text":1455},[],"Hugo Horlings has spent years looking at things too small for the human eye to see clearly. He is also someone who thinks the most important next step in medicine is not a better algorithm. It is a better relationship between the people who build the tools and the people who use them.",[],{"_key":1458,"_type":502},"336218f43029",{"_key":1460,"_type":443,"text":1461},"67fdbe622f88",[1462,1471,1483,1495,1507],{"_key":1463,"_type":130,"children":1464,"markDefs":1469,"style":1470},"220dd7c6637a",[1465],{"_key":1466,"_type":134,"marks":1467,"text":1468},"0d0d0f1e322c",[684],"Hugo hopes this conversation multiplies. For those already working with AI, or just thinking about getting started, he shared a few things he has learned from doing it every day.",[],"h4",{"_key":1472,"_type":130,"children":1473,"markDefs":1482,"style":138},"7073d3b7ba51",[1474,1478],{"_key":1475,"_type":134,"marks":1476,"text":1477},"8e5a900014f6",[468],"Start small:",{"_key":1479,"_type":134,"marks":1480,"text":1481},"f4231bd28a91",[],"\nDon't look for the grand application. Find the most frustrating part in your daily workflow and bring it to AI first. And before trying anything clinical, start at home. A subscription to one of the large language models, ChatGPT, Claude, or Gemini, is enough to begin. ",[],{"_key":1484,"_type":130,"children":1485,"markDefs":1494,"style":138},"2726cc2014a9",[1486,1490],{"_key":1487,"_type":134,"marks":1488,"text":1489},"32aaf96a4227",[468],"Becoming a more effective decision-maker:",{"_key":1491,"_type":134,"marks":1492,"text":1493},"f33516cd1c9a",[],"\nHere is something that does not get said enough about AI in clinical practice: it does not reduce the number of decisions you make; it changes what those decisions are. AI handles the searching, the summarising, the reconstruction of fragmented records. What that frees up is not purely about reducing work. It is about freeing up space for the work that matters most. You move from information gatherer to full-time decision-maker, and that shift takes getting used to.",[],{"_key":1496,"_type":130,"children":1497,"markDefs":1506,"style":138},"4e82a023ed37",[1498,1502],{"_key":1499,"_type":134,"marks":1500,"text":1501},"13a9c964cb4b",[468],"Prompt with precision:",{"_key":1503,"_type":134,"marks":1504,"text":1505},"2fd00c65bb43",[],"\nVague prompts produce vague answers. Give the model a role and a context: instead of asking a broad question, tell it who it is and what you need. Ask it to verify its own output before giving you the answer. And if you want something in a specific format, paste in a blank template or a fully anonymised example. The model will follow the structure.",[],{"_key":1508,"_type":130,"children":1509,"markDefs":1518,"style":138},"06005e3126b9",[1510,1514],{"_key":1511,"_type":134,"marks":1512,"text":1513},"e1501423ae7d",[468],"Don't build alone:",{"_key":1515,"_type":134,"marks":1516,"text":1517},"8f413385215c",[],"\nThis applies to individual clinicians and to hospitals. Neither can do it well in isolation. The results come from working alongside people with complementary knowledge, whether that is a technically-minded colleague or an external partner that can help build the right solutions and infrastructure. One of those examples is kaiko, built specifically for clinical environments and already live at the Netherlands Cancer Institute.",[],{"_key":1520,"_type":1071,"references":1521},"200040020c79",[1522,1523,1524,1525,1526],"OPTImaL:Optimisation of Treatment for Patients With Low Stage Triple-negative Breast Cancer With High Stromal Tumor-infiltrating Lymphocytes (OPTImaL), The Netherlands Cancer Institute https://clinicaltrials.gov/study/NCT06476119","Geurts et al. (2024). Tumor-infiltrating lymphocytes in patients with stage I triple-negative breast cancer untreated with chemotherapy. JAMA Oncology, 10(8), 1077–1086. https://jamanetwork.com/journals/jamaoncology/fullarticle/2820527","https://www.avl.nl/alles-over-kanker/informatie-over-klinische-studies-trials/trials/borstkanker-optimal-studie/","https://www.webpathology.com/images/dermpath/proliferations-and-neoplasms/melanoma---general/40630","https://flightsafety.org/asw-article/making-a-list/",{"_key":1528,"_type":443,"text":1529},"7f9342d589a3",[1530],{"_key":1531,"_type":130,"children":1532,"markDefs":1537,"style":1538},"e3c68de45dfb",[1533],{"_key":1534,"_type":134,"marks":1535,"text":1536},"8728db829842",[],"Our systems are currently developed and tested in research and validation settings. Some of our products focus on supporting administrative and workflow processes and are not medical devices. Any future clinical decision support use would be subject to regulatory approval and compliance with applicable medical device frameworks, including the EU Medical Device Regulation (MDR).",[],"disclaimer","2026-06-25T09:00:00.000Z",{"_type":89,"asset":1541},{"_ref":1542,"_type":92},"image-fc02314c846ed2051870f41a78c1e69b33ec829b-3017x1412-png",[1544,1547],{"_key":1545,"_ref":1546,"_type":92},"0158751a0ab9","3877c07b-5faa-4f0a-8653-a9440b5fb289",{"_key":1548,"_ref":1549,"_type":92},"e9eda70d9881","6bbbd693-3a09-4061-972d-b72307863bf3",{"_type":1087,"current":1551},"AI-in-Pathology-Hugo-Horlings","A conversation with Hugo Horlings, pathologist and researcher at the Netherlands Cancer Institute.","The nudge to look again, clinical AI as told by a pathologist",[1555,1558,1561],{"_key":1556,"_ref":1557,"_type":92},"ad265e09e351","87071b12-f012-4898-9928-e78a41e6f88e",{"_key":1559,"_ref":1560,"_type":92},"bcb29e6856e2","c723429d-eae1-42ad-bbd1-01e130a227ae",{"_key":1562,"_ref":1563,"_type":92},"a9901122ba55","c8c35a3a-1afb-4866-ae93-e5722f2f845a",{"_createdAt":1565,"_id":1549,"_rev":1566,"_system":1567,"_type":428,"_updatedAt":1570,"area":1571,"categories":1572,"content":1576,"date":1732,"featuredImage":1733,"hidePostMeta":1736,"isHidden":94,"slug":1737,"title":1739,"topics":1740},"2026-05-06T11:48:49Z","ZL7oNypv5EIL16VHo5gHKx",{"base":1568},{"id":1549,"rev":1569},"RrQSLznHnmNiNzNAQDrRpV","2026-06-24T17:42:08Z",[1112],[1573],{"_key":1574,"_ref":1575,"_type":92},"eb6985b99422","4d4cf8a5-8d65-4174-96ac-b36f31bed9cc",[1577,1596,1601,1603,1695,1697],{"_key":1578,"_type":443,"text":1579},"194cdd717cf6",[1580,1588],{"_key":1581,"_type":130,"children":1582,"markDefs":1587,"style":453},"b8ead4938063",[1583],{"_key":1584,"_type":134,"marks":1585,"text":1586},"ff2262fbd7ca",[],"Rewatch the session ",[],{"_key":1589,"_type":130,"children":1590,"markDefs":1595,"style":138},"f51fe3affbcd",[1591],{"_key":1592,"_type":134,"marks":1593,"text":1594},"39f444d51f94",[],"The live-stream has ended, but you can now directly rewatch the full session below, without registration.",[],{"_key":1597,"_type":1598,"title":1599,"videoUrl":1600},"a995a72fdd11","liveVideoSection"," ","https://www.youtube.com/watch?v=1x2VR2YS_zM",{"_key":1602,"_type":502},"e370c364b26c",{"_key":1604,"_type":443,"text":1605},"c71a5e176744",[1606,1614,1622,1633,1644,1655,1666,1674,1681,1688],{"_key":1607,"_type":130,"children":1608,"markDefs":1613,"style":453},"ad67135dbd9a",[1609],{"_key":1610,"_type":134,"marks":1611,"text":1612},"89675ddcb102",[],"Key moments",[],{"_key":1615,"_type":130,"children":1616,"markDefs":1621,"style":138},"cce648631373",[1617],{"_key":1618,"_type":134,"marks":1619,"text":1620},"57e16cbf0dad",[],"What you can expect in today’s session, timestamped:",[],{"_key":1623,"_type":130,"children":1624,"level":101,"listItem":474,"markDefs":1632,"style":138},"75959a0d4326",[1625,1628],{"_key":1618,"_type":134,"marks":1626,"text":1627},[468],"11:00 - 11:20:",{"_key":1629,"_type":134,"marks":1630,"text":1631},"cce920c90aeb",[]," Panel discussion with Maurice van den Bosch, Annemarie van 't Veen, and Tim Jansen",[],{"_key":1634,"_type":130,"children":1635,"level":101,"listItem":474,"markDefs":1643,"style":138},"676f82a39b9f",[1636,1639],{"_key":1618,"_type":134,"marks":1637,"text":1638},[468],"11:20 - 11:22:",{"_key":1640,"_type":134,"marks":1641,"text":1642},"6701d7d3baa2",[]," Launching kaiko: a word from our founder Thomas Hufener and team",[],{"_key":1645,"_type":130,"children":1646,"level":101,"listItem":474,"markDefs":1654,"style":138},"7952aaee4a4e",[1647,1650],{"_key":1618,"_type":134,"marks":1648,"text":1649},[468],"11:22 - 11:29:",{"_key":1651,"_type":134,"marks":1652,"text":1653},"6074c2aa54ed",[]," Demo: kaiko, the clinical AI workspace",[],{"_key":1656,"_type":130,"children":1657,"level":101,"listItem":474,"markDefs":1665,"style":138},"fbb88d6cf721",[1658,1661],{"_key":1618,"_type":134,"marks":1659,"text":1660},[468],"11:29 -11:40:",{"_key":1662,"_type":134,"marks":1663,"text":1664},"7b6d72b29608",[]," Fireside chat with Surgical Oncologist Michel Wouters",[],{"_key":1667,"_type":130,"children":1668,"markDefs":1673,"style":138},"cf6d1682e74d",[1669],{"_key":1670,"_type":134,"marks":1671,"text":1672},"ec829b6ffa8a",[],"",[],{"_key":1675,"_type":130,"children":1676,"markDefs":1680,"style":138},"8cff17665ff6",[1677],{"_key":1678,"_type":134,"marks":1679,"text":1672},"4a5efdadf4af",[],[],{"_key":1682,"_type":130,"children":1683,"markDefs":1687,"style":138},"de80a5b63438",[1684],{"_key":1685,"_type":134,"marks":1686,"text":1672},"802487ba0f21",[],[],{"_key":1689,"_type":130,"children":1690,"markDefs":1694,"style":138},"72b76a4c74e2",[1691],{"_key":1692,"_type":134,"marks":1693,"text":1672},"9416bbe817cb",[],[],{"_key":1696,"_type":502},"fc00a724ba19",{"_key":1698,"_type":443,"text":1699},"d77de852d77e",[1700,1708,1716,1724],{"_key":1701,"_type":130,"children":1702,"markDefs":1707,"style":453},"ab979e0f9a1f",[1703],{"_key":1704,"_type":134,"marks":1705,"text":1706},"6d81fb291843",[],"About the session",[],{"_key":1709,"_type":130,"children":1710,"markDefs":1715,"style":138},"dd0cfcd345e8",[1711],{"_key":1712,"_type":134,"marks":1713,"text":1714},"789973b592a8",[],"Hospitals are investing in AI, building strategies, and running pilots. What most of them haven't cracked yet is how to make it last.",[],{"_key":1717,"_type":130,"children":1718,"markDefs":1723,"style":138},"032c19da66a9",[1719],{"_key":1720,"_type":134,"marks":1721,"text":1722},"d8f78abf5e03",[],"On June 24, 2026, we explored why, and what the hospitals that figured it out did differently. A CEO, an AI accelerator lead, and a critical care specialist from three Dutch hospitals joined Wouter Gude, author of the AI Monitor for Hospitals 2026, to share the decisions, the resistance, and the lessons they wish they had earlier.\n\nYou will also see kaiko, the clinical AI workspace, live and already running at leading European hospitals, preparing a patient consultation. And you will hear from a surgeon at NKI-AVL on what it changed for him and his patients.",[],{"_key":1725,"_type":130,"children":1726,"markDefs":1731,"style":138},"cbbcde6ae15e",[1727],{"_key":1728,"_type":134,"marks":1729,"text":1730},"7b9c147d605b",[],"Walk away with a clear answer to the question most hospital leaders are still sitting with: what does it actually take?",[],"2026-06-24T09:14:00.000Z",{"_type":89,"asset":1734},{"_ref":1735,"_type":92},"image-73ddc07037b8c5bd13322caf6b491eb99da12110-2010x1140-png",true,{"_type":1087,"current":1738},"kaiko-launch-watch","Pilot limbo: why most clinical AI never scales, and how some hospitals break the cycle  ",[1741,1743,1746],{"_key":1742,"_ref":1096,"_type":92},"e2f0af87f364",{"_key":1744,"_ref":1745,"_type":92},"6f8f2ef3b7c4","8b094fc2-ff86-45f9-a1ee-8bd85a2cdbe5",{"_key":1747,"_ref":1560,"_type":92},"fd620845dded",{"_createdAt":1749,"_id":1085,"_rev":1750,"_system":1751,"_type":428,"_updatedAt":1754,"area":1755,"authors":1756,"categories":1760,"content":1763,"date":2433,"featuredImage":2434,"isHidden":94,"relatedPosts":2437,"slug":2441,"subHeading":2443,"title":2444,"topics":2445},"2026-06-02T11:36:58Z","DBg3LM0vF22ywEKWDVO0dm",{"base":1752},{"id":1085,"rev":1753},"GInI3CC6S35oK4J4gwnIB1","2026-06-10T09:23:43Z",[431],[1757],{"_key":1758,"_ref":1759,"_type":92},"6b51935126c4","737d8376-7ded-403e-98a6-30fa6b4c87e7",[1761],{"_key":1762,"_ref":439,"_type":92},"a78b4a06e8fc",[1764,1770,1853,1855,1877,1883,1922,1924,1952,1958,1969,1975,2022,2041,2047,2049,2076,2087,2093,2120,2175,2177,2248,2250,2406,2408],{"_key":1765,"_type":626,"caption":1766,"image":1767},"9566649717a4","Figure 1-4: Ground-truth annotations vs CoralBay predictions on various datasets.",{"_type":89,"asset":1768},{"_ref":1769,"_type":92},"image-f2f8cd6857b5225f5e7a438ea0fbd3b8520ccbbb-1280x720-gif",{"_key":1771,"_type":443,"text":1772},"64220c0acaa7",[1773,1781,1797,1805,1817,1829,1841],{"_key":1774,"_type":130,"children":1775,"markDefs":1780,"style":138},"35765f76c825",[1776],{"_key":1777,"_type":134,"marks":1778,"text":1779},"8c81e04012ee",[],"Each year, over 300 million CT scans are captured as rich, high-fidelity 3D volumes that give doctors a detailed view of the human body for diagnosis and surgical planning. However, the status quo in current AI models is to process these volumes as stacks of 2D slices, effectively flattening the data and discarding critical spatial relationships, or to rely on complicated, highly specialised methods that are difficult to scale and build upon.",[],{"_key":1782,"_type":130,"children":1783,"markDefs":1796,"style":138},"b331bc6804ad",[1784,1788,1792],{"_key":1785,"_type":134,"marks":1786,"text":1787},"7bca31b804f0",[],"Enter ",{"_key":1789,"_type":134,"marks":1790,"text":1791},"cd673be05ab4",[468],"CoralBay",{"_key":1793,"_type":134,"marks":1794,"text":1795},"aacb86bcdcce",[],": our native 3D foundation model for CT, pre-trained on only 11K unlabeled volumes. With a simple architecture and training pipeline, it produces meaningful representations from raw data that generalize across the full clinical stack, from scan-level classification down to fine-grained lesion segmentation.",[],{"_key":1798,"_type":130,"children":1799,"markDefs":1804,"style":453},"54b86dc85d82",[1800],{"_key":1801,"_type":134,"marks":1802,"text":1803},"6a2cc7267166",[468],"Quick Highlights",[],{"_key":1806,"_type":130,"children":1807,"level":101,"listItem":474,"markDefs":1816,"style":138},"4d251260c971",[1808,1812],{"_key":1809,"_type":134,"marks":1810,"text":1811},"05b0b7dc5891",[468],"Native 3D Intelligence:",{"_key":1813,"_type":134,"marks":1814,"text":1815},"e07ecffd95d4",[]," Understands anatomy as continuous volumetric structures.",[],{"_key":1818,"_type":130,"children":1819,"level":101,"listItem":474,"markDefs":1828,"style":138},"8caec661b008",[1820,1824],{"_key":1821,"_type":134,"marks":1822,"text":1823},"abe2ae64d09d",[468],"Simple, Strong, Generalizable:",{"_key":1825,"_type":134,"marks":1826,"text":1827},"1b030a7961bb",[]," Strong multi-task performance from a simple training pipeline and just 11K unlabeled volumes, ready to adapt to various downstream tasks.",[],{"_key":1830,"_type":130,"children":1831,"level":101,"listItem":474,"markDefs":1840,"style":138},"fee963bbe5c7",[1832,1836],{"_key":1833,"_type":134,"marks":1834,"text":1835},"18b325e8ae18",[468],"Clinical Depth:",{"_key":1837,"_type":134,"marks":1838,"text":1839},"55efa3c6f50e",[]," A single backbone with built-in invariance to HU values, excelling across the full task spectrum from broad organ identification to fine-grained tumour segmentation",[],{"_key":1842,"_type":130,"children":1843,"level":101,"listItem":474,"markDefs":1852,"style":138},"43c283f1419c",[1844,1848],{"_key":1845,"_type":134,"marks":1846,"text":1847},"21d4328d42dc",[468],"Open Science:",{"_key":1849,"_type":134,"marks":1850,"text":1851},"bfdcf0b49a91",[]," Fully open-source weights and benchmarks to accelerate global medical AI research.",[],{"_key":1854,"_type":502},"4be33ed573b9",{"_key":1856,"_type":443,"text":1857},"f8ff5b1d9a48",[1858,1866],{"_key":1859,"_type":130,"children":1860,"markDefs":1865,"style":453},"bd0845d89d25",[1861],{"_key":1862,"_type":134,"marks":1863,"text":1864},"5bd09ab7d8c7",[468],"The Challenge: 3D is Different",[],{"_key":1867,"_type":130,"children":1868,"markDefs":1876,"style":138},"b8528c9d14ad",[1869,1873],{"_key":1870,"_type":134,"marks":1871,"text":1872},"1576c6dfdc1d",[],"Standard vision models assume 2D RGB images; three colour channels, consistent resolution, and millions of available labeled examples. CT scans violate every one of these assumptions.",{"_key":1862,"_type":134,"marks":1874,"text":1875},[468],"\n",[],{"_key":1878,"_type":626,"caption":1879,"image":1880},"b34e97709062","Figure 5: Challenges in 3D data representation. Left: Narrow windows clarify soft tissue; wide windows preserve high-density data. Center: Thin slices increase resolution; thick slices improve SNR but cause blurring. Right: 3D consistency is required across all orthogonal planes.",{"_type":89,"asset":1881},{"_ref":1882,"_type":92},"image-b15c3a2413e2743044aa8a6f40419291816065e1-2857x860-png",{"_key":1884,"_type":443,"text":1885},"7170efbff7f2",[1886,1898,1910],{"_key":1887,"_type":130,"children":1888,"level":101,"listItem":474,"markDefs":1897,"style":138},"13c3b67b8f91",[1889,1893],{"_key":1890,"_type":134,"marks":1891,"text":1892},"951db6b47c8a",[468],"Hounsfield Units, not pixels",{"_key":1894,"_type":134,"marks":1895,"text":1896},"68968e710212",[],": Intensities reflect tissue density, not color. Different window settings highlight different anatomy (e.g., lungs vs. soft tissue), so models must be robust to multiple visualizations of the same scan.",[],{"_key":1899,"_type":130,"children":1900,"level":101,"listItem":474,"markDefs":1909,"style":138},"d5c75cd01feb",[1901,1905],{"_key":1902,"_type":134,"marks":1903,"text":1904},"fbec56375347",[468],"Anisotropic resolution",{"_key":1906,"_type":134,"marks":1907,"text":1908},"48c80c0a9f45",[],": Slice thickness varies (e.g., 1–5 mm), causing partial volume effects. Thicker slices improve signal-to-noise but blur fine structures, unlike typical image noise.",[],{"_key":1911,"_type":130,"children":1912,"level":101,"listItem":474,"markDefs":1921,"style":138},"a5fd44829f4a",[1913,1917],{"_key":1914,"_type":134,"marks":1915,"text":1916},"88bb51682b5d",[468],"Volumetric spatial complexity",{"_key":1918,"_type":134,"marks":1919,"text":1920},"e4263fe818ca",[],": CT is inherently 3D. Small findings, including early-stage tumours that may occupy only a handful of voxels, require integrating information across the full 3D space to be reliably detected. Treating slices independently breaks the spatial continuity that makes these critical, hard-to-spot findings visible in the first place.",[],{"_key":1923,"_type":502},"17ece86911bc",{"_key":1925,"_type":443,"text":1926},"910b8f205540",[1927,1935,1944],{"_key":1928,"_type":130,"children":1929,"markDefs":1934,"style":453},"e1031f7eaa2b",[1930],{"_key":1931,"_type":134,"marks":1932,"text":1933},"1958c0300072",[468],"The CoralBay Approach",[],{"_key":1936,"_type":130,"children":1937,"markDefs":1942,"style":1943},"24cf30dce250",[1938],{"_key":1939,"_type":134,"marks":1940,"text":1941},"cfef1df31105",[468],"Hierarchical 3D Feature Learning",[],"h3",{"_key":1945,"_type":130,"children":1946,"markDefs":1951,"style":138},"22efd7486e59",[1947],{"_key":1948,"_type":134,"marks":1949,"text":1950},"26123db28de3",[],"CoralBay extends DINO self-supervised learning to native 3D CT volumes. Using a hierarchical 3D Swin-Transformer, it captures long-range spatial relationships efficiently and learns multi-scale features ranging from organ-level anatomy to fine structures such as vessels.",[],{"_key":1953,"_type":626,"caption":1954,"image":1955},"fc44b831cb37","Figure 6: CoralBay’s technical overviews of the training pipeline",{"_type":89,"asset":1956},{"_ref":1957,"_type":92},"image-dbb59b4f247bc0d580e639560876d93759308654-2983x830-png",{"_key":1959,"_type":443,"text":1960},"e24b6ad5a981",[1961],{"_key":1962,"_type":130,"children":1963,"markDefs":1968,"style":1943},"b5b27746eff8",[1964],{"_key":1965,"_type":134,"marks":1966,"text":1967},"17448af3114b",[468],"Radiology-Specific Augmentations",[],{"_key":1970,"_type":626,"caption":1971,"image":1972},"d5b91e32677e","Figure 7: HU ranges for pre-training data augmentation.",{"_type":89,"asset":1973},{"_ref":1974,"_type":92},"image-81151700e6702f95aea34f954a36249de02bdf7b-1871x662-png",{"_key":1976,"_type":443,"text":1977},"cef125988b64",[1978,1986,1998,2010],{"_key":1979,"_type":130,"children":1980,"markDefs":1985,"style":138},"0183fefb6cf8",[1981],{"_key":1982,"_type":134,"marks":1983,"text":1984},"0aebd20a500b",[],"CoralBay replaces generic image augmentations with CT-specific transformations that reflect real-world imaging variability.",[],{"_key":1987,"_type":130,"children":1988,"level":101,"listItem":474,"markDefs":1997,"style":138},"6cb032b32dc8",[1989,1993],{"_key":1990,"_type":134,"marks":1991,"text":1992},"ed6a8d78c81c",[468],"Random HU Windowing:",{"_key":1994,"_type":134,"marks":1995,"text":1996},"4950f307840f",[]," Samples from multiple clinically relevant HU ranges (e.g., lung, liver, brain, abdomen, and full CT) to learn features that are robust across viewing settings.",[],{"_key":1999,"_type":130,"children":2000,"level":101,"listItem":474,"markDefs":2009,"style":138},"ce21392c405f",[2001,2005],{"_key":2002,"_type":134,"marks":2003,"text":2004},"4773b4688429",[468],"Scanner robustness:",{"_key":2006,"_type":134,"marks":2007,"text":2008},"05f3208926e3",[]," Gaussian smoothing and histogram shifts simulate differences in scanners and reconstruction protocols, improving generalization across sites.",[],{"_key":2011,"_type":130,"children":2012,"level":101,"listItem":474,"markDefs":2021,"style":138},"2acdfbb129d2",[2013,2017],{"_key":2014,"_type":134,"marks":2015,"text":2016},"cfec638253be",[468],"Local-to-global 3D context:",{"_key":2018,"_type":134,"marks":2019,"text":2020},"e6d600c2cfe9",[]," Combines large and small 3D crops during training, helping the model link fine pathological details with their surrounding anatomy.\n\n",[],{"_key":2023,"_type":443,"text":2024},"15ad5094a40b",[2025,2033],{"_key":2026,"_type":130,"children":2027,"markDefs":2032,"style":1943},"be04f790ac1e",[2028],{"_key":2029,"_type":134,"marks":2030,"text":2031},"8da62902f193",[468],"Scan-Level Inference via Sliding Window",[],{"_key":2034,"_type":130,"children":2035,"markDefs":2040,"style":138},"b01bb1b885eb",[2036],{"_key":2037,"_type":134,"marks":2038,"text":2039},"2e22e58a177a",[],"While the backbone is trained on 96×96×96 crops, inference uses a sliding-window approach: the full scan is divided into overlapping 3D patches, each encoded independently, then stitched together. For classification, pooling merges the patch-level features into a single scan-level vector. For segmentation, features are passed to a Swin-UNETR decoder with skip connections, which preserves fine boundary details for voxel-wise segmentation.",[],{"_key":2042,"_type":626,"caption":2043,"image":2044},"b53c3ec1a4f0","Figure 8a-b: 3D scan-level classification (left) and segmentation (right) sliding window Inference",{"_type":89,"asset":2045},{"_ref":2046,"_type":92},"image-9c7d3b91caa94b4ccf1f486106b39d3adff64310-1080x608-gif",{"_key":2048,"_type":502},"9d1f2386f430",{"_key":2050,"_type":443,"text":2051},"373492d8a965",[2052,2060,2068],{"_key":2053,"_type":130,"children":2054,"markDefs":2059,"style":453},"27742a556a73",[2055],{"_key":2056,"_type":134,"marks":2057,"text":2058},"49a15d39b191",[468],"How CoralBay Compares",[],{"_key":2061,"_type":130,"children":2062,"markDefs":2067,"style":138},"d9730d5a19bc",[2063],{"_key":2064,"_type":134,"marks":2065,"text":2066},"5ef80e386a2f",[],"As the CT foundation model space matures, CoralBay can be understood along three axes: how well it transfers to downstream tasks with both frozen and fine-tuned encoders, how little pre-training data it requires, and how broadly it generalises across classification and segmentation tasks.",[],{"_key":2069,"_type":130,"children":2070,"markDefs":2075,"style":138},"edf6185361e2",[2071],{"_key":2072,"_type":134,"marks":2073,"text":2074},"9a9870265e97",[],"To test this, we evaluated CoralBay on 11 datasets spanning classification and segmentation across diverse anatomical targets, with two model variants: CoralBayU96B (53.2M) and CoralBayU96H (847M).",[],{"_key":2077,"_type":443,"text":2078},"546419008d33",[2079],{"_key":2080,"_type":130,"children":2081,"markDefs":2086,"style":453},"454658f66a2e",[2082],{"_key":2083,"_type":134,"marks":2084,"text":2085},"000230fac4c6",[],"Quantitative Results",[],{"_key":2088,"_type":626,"caption":2089,"image":2090},"7a5cce16e0ad","Table A: Quantitative performance across classification (Multi-class Accuracy/Binary AUROC) and segmentation (Dice score) tasks, as evaluated via the eva framework.",{"_type":89,"asset":2091},{"_ref":2092,"_type":92},"image-d3bd456a4f8a0285bf1592543ff451c5d2655910-3881x1864-png",{"_key":2094,"_type":443,"text":2095},"f0202c030fc1",[2096,2108],{"_key":2097,"_type":130,"children":2098,"markDefs":2107,"style":138},"444f6936a4f7",[2099,2103],{"_key":2100,"_type":134,"marks":2101,"text":2102},"0a2fa2c6c444",[468],"- Classification:",{"_key":2104,"_type":134,"marks":2105,"text":2106},"246c3c863c32",[]," CoralBayU96H achieves the best or tied-best frozen-encoder results across all four classification benchmarks — organ identification (OrganMNIST3D), lung nodule malignancy (NoduleMNIST3D, LUNA25), and COVID-19 classification (CC-CCII).",[],{"_key":2109,"_type":130,"children":2110,"markDefs":2119,"style":138},"72429b2441da",[2111,2115],{"_key":2112,"_type":134,"marks":2113,"text":2114},"65e6b282dfa5",[468],"- Segmentation:",{"_key":2116,"_type":134,"marks":2117,"text":2118},"66e3b870c73b",[]," With a frozen encoder and a lightweight 22.8M-parameter decoder, CoralBay performs comparably to VoCo across seven segmentation benchmarks despite the significant data gap.\n\n",[],{"_key":2121,"_type":443,"text":2122},"11e22cc37247",[2123,2131,2139,2151,2163],{"_key":2124,"_type":130,"children":2125,"markDefs":2130,"style":453},"038664d34317",[2126],{"_key":2127,"_type":134,"marks":2128,"text":2129},"a88282b576bf",[],"Qualitative Analysis",[],{"_key":2132,"_type":130,"children":2133,"markDefs":2138,"style":138},"6e4b26c4cad3",[2134],{"_key":2135,"_type":134,"marks":2136,"text":2137},"eb8962ac6119",[],"Visualising the model's performance (Figures 1-4) reveals its precision across diverse anatomical challenges:",[],{"_key":2140,"_type":130,"children":2141,"level":101,"listItem":474,"markDefs":2150,"style":138},"ff3b8ca7829f",[2142,2146],{"_key":2143,"_type":134,"marks":2144,"text":2145},"bb4643edcecb",[468],"Multi-Organ Segmentation:",{"_key":2147,"_type":134,"marks":2148,"text":2149},"9b40d7d71e77",[]," On benchmarks like BTCV and FLARE22, the model accurately delineates complex abdominal structures including the liver, spleen, and kidneys with high spatial consistency.",[],{"_key":2152,"_type":130,"children":2153,"level":101,"listItem":474,"markDefs":2162,"style":138},"24bc2f23c9f5",[2154,2158],{"_key":2155,"_type":134,"marks":2156,"text":2157},"b15a6cd2934f",[468],"Lesion and tumour boundaries:",{"_key":2159,"_type":134,"marks":2160,"text":2161},"b5f90a41d68e",[]," Evaluation on LiTS17 and MSD Task 7 demonstrates its ability to resolve fine-grained tumor boundaries within the liver and pancreas, critical for clinical diagnostics.",[],{"_key":2164,"_type":130,"children":2165,"level":101,"listItem":474,"markDefs":2174,"style":138},"e824eaf2088d",[2166,2170],{"_key":2167,"_type":134,"marks":2168,"text":2169},"7101abc76ded",[468],"Reliability:",{"_key":2171,"_type":134,"marks":2172,"text":2173},"93edb776f4ca",[]," The difference maps highlight minimal variance between CoralBay segmentations and ground-truth annotations, confirming robust feature capture in low-contrast regions.",[],{"_key":2176,"_type":502},"98cc8424d8fb",{"_key":2178,"_type":443,"text":2179},"28f603846357",[2180,2188,2196,2216,2236],{"_key":2181,"_type":130,"children":2182,"markDefs":2187,"style":453},"2958a8aa097b",[2183],{"_key":2184,"_type":134,"marks":2185,"text":2186},"5eb56301d4d0",[468],"What the Ablations Reveal",[],{"_key":2189,"_type":130,"children":2190,"markDefs":2195,"style":138},"624b97cc39b6",[2191],{"_key":2192,"_type":134,"marks":2193,"text":2194},"e79ff7601a16",[],"Ablation studies in the paper confirm that CoralBay’s performance arises from a combination of structural and data-driven innovations:",[],{"_key":2197,"_type":130,"children":2198,"level":101,"listItem":474,"markDefs":2215,"style":138},"9786e5156e10",[2199,2203,2207,2211],{"_key":2200,"_type":134,"marks":2201,"text":2202},"aabe76e00128",[468],"Native 3D Dominance:",{"_key":2204,"_type":134,"marks":2205,"text":2206},"adce4cd9683d",[]," Switching from the native dino 2D to CoralBay 3D spatial modeling provides a substantial improvement of ",{"_key":2208,"_type":134,"marks":2209,"text":2210},"12e0269e2fba",[468],"20%",{"_key":2212,"_type":134,"marks":2213,"text":2214},"3a46a6f7539b",[]," in Dice score (BTCV), proving that consistent 3D inductive biases are essential for medical volumes.",[],{"_key":2217,"_type":130,"children":2218,"level":101,"listItem":474,"markDefs":2235,"style":138},"fe64aed92a06",[2219,2223,2227,2231],{"_key":2220,"_type":134,"marks":2221,"text":2222},"e7647ed0f39d",[468],"Effective Scaling",{"_key":2224,"_type":134,"marks":2225,"text":2226},"195dd081c52b",[],": Segmentation accuracy improves consistently as both the model size and the pre-training dataset grow, resulting in an overall ",{"_key":2228,"_type":134,"marks":2229,"text":2230},"80af9713004a",[468],"15% improvement",{"_key":2232,"_type":134,"marks":2233,"text":2234},"5d5a95f56819",[]," (LiTS17). This demonstrates the framework’s ability to benefit from larger data pools and suggests that training with even more data could further strengthen the model.",[],{"_key":2237,"_type":130,"children":2238,"level":101,"listItem":474,"markDefs":2247,"style":138},"bdce70c68041",[2239,2243],{"_key":2240,"_type":134,"marks":2241,"text":2242},"eefc607d4e17",[468],"Superior Label Efficiency:",{"_key":2244,"_type":134,"marks":2245,"text":2246},"5b3b36cdcf4e",[]," On challenging tumor tasks, self-supervised pre-training acts as a powerful prior, outperforming heavily tuned models in low-data settings.",[],{"_key":2249,"_type":502},"8c07382648bb",{"_key":2251,"_type":443,"text":2252},"fcdb71887c26",[2253,2261,2295,2302,2310,2318,2330,2353,2365],{"_key":2254,"_type":130,"children":2255,"markDefs":2260,"style":453},"6da8f2643ea9",[2256],{"_key":2257,"_type":134,"marks":2258,"text":2259},"509fc72c3ffe",[468],"Why this matters for kaiko",[],{"_key":2262,"_type":130,"children":2263,"markDefs":2294,"style":138},"39467400a5c4",[2264,2267,2271,2274,2278,2282,2286,2290],{"_key":2265,"_type":134,"marks":2266,"text":1791},"c0638d99be1f",[468],{"_key":2268,"_type":134,"marks":2269,"text":2270},"f43458ba626d",[]," establishes a powerful 3D foundation for radiological data. Today, it masters native spatial reasoning and physics-informed intensity invariance, delivering highly data-efficient, clinically robust models for classification and segmentation across organs, pathologies, and scanners. This raises the ceiling for high‑resolution pathology detection while also providing a standard, open benchmark through the 3D radiology leaderboard. \n\nLooking ahead, ",{"_key":2272,"_type":134,"marks":2273,"text":1791},"941e381f5bd5",[468],{"_key":2275,"_type":134,"marks":2276,"text":2277},"4614f3441733",[]," serves as a visual anchor for ",{"_key":2279,"_type":134,"marks":2280,"text":2281},"23837e082ca5",[468],"Multimodal Medical Intelligence",{"_key":2283,"_type":134,"marks":2284,"text":2285},"6b0544eff428",[],", where AI systems jointly interpret imaging, text, lab values, and other clinical signals in a unified clinical context. Built on a strong 3D visual backbone, it moves toward an ",{"_key":2287,"_type":134,"marks":2288,"text":2289},"641b143a9e24",[468],"agentic vision paradigm",{"_key":2291,"_type":134,"marks":2292,"text":2293},"469ed812d54a",[],", where models go beyond interpretation to actively reason over scans, plan multi-step analyses, and coordinate tool use across modalities and time.",[],{"_key":2296,"_type":130,"children":2297,"markDefs":2301,"style":138},"f884f3cd83a7",[2298],{"_key":2299,"_type":134,"marks":2300,"text":1672},"71631a68b7ae",[],[],{"_key":2303,"_type":130,"children":2304,"markDefs":2309,"style":453},"f17bcd4148c4",[2305],{"_key":2306,"_type":134,"marks":2307,"text":2308},"3506c3bc6681",[468],"Open Science",[],{"_key":2311,"_type":130,"children":2312,"markDefs":2317,"style":138},"8ad16fdfcb89",[2313],{"_key":2314,"_type":134,"marks":2315,"text":2316},"b2767c4e8ed4",[],"CoralBay promotes open, reproducible evaluation in medical AI through core contributions:",[],{"_key":2319,"_type":130,"children":2320,"level":101,"listItem":474,"markDefs":2329,"style":138},"e8a7a8dcb4c7",[2321,2325],{"_key":2322,"_type":134,"marks":2323,"text":2324},"1aa0531f239e",[468],"Comprehensive Whitepaper:",{"_key":2326,"_type":134,"marks":2327,"text":2328},"ae32bdf52146",[]," We provide a detailed whitepaper describing the model, training process, datasets, evaluation setup, and results.",[],{"_key":2331,"_type":130,"children":2332,"level":101,"listItem":474,"markDefs":2350,"style":138},"f4b10baeca14",[2333,2337,2341,2346],{"_key":2334,"_type":134,"marks":2335,"text":2336},"11e7b8fa0e9e",[468],"Standardized Benchmarking:",{"_key":2338,"_type":134,"marks":2339,"text":2340},"17b681ae0a2b",[]," We’re scaling 🔗 ",{"_key":2342,"_type":134,"marks":2343,"text":2345},"556e5f7ffd40",[468,2344],"ca96badbe275","eva",{"_key":2347,"_type":134,"marks":2348,"text":2349},"14d64945d831",[]," — our open-source evaluation framework — to include comprehensive support for 3D radiology. This includes standardized data loaders, model backbones, and a public leaderboard designed to make 3D medical AI research transparent, reproducible, and easily comparable.",[2351],{"_key":2344,"_type":622,"href":2352},"https://github.com/kaiko-ai/eva",{"_key":2354,"_type":130,"children":2355,"level":101,"listItem":474,"markDefs":2364,"style":138},"9057a5c47442",[2356,2360],{"_key":2357,"_type":134,"marks":2358,"text":2359},"71e126998f48",[468],"Open Access:",{"_key":2361,"_type":134,"marks":2362,"text":2363},"f3b605346cff",[]," Model weights are available on GitHub and Hugging Face.",[],{"_key":2366,"_type":130,"children":2367,"markDefs":2399,"style":138},"402b38cf12a7",[2368,2372,2377,2381,2386,2390,2395],{"_key":2369,"_type":134,"marks":2370,"text":2371},"68c6ea89183c",[],"Explore the 🔗 ",{"_key":2373,"_type":134,"marks":2374,"text":2376},"390e9ca438c4",[2375],"3ca4e837b5ae","paper",{"_key":2378,"_type":134,"marks":2379,"text":2380},"f9ffddb50552",[],", 🔗 ",{"_key":2382,"_type":134,"marks":2383,"text":2385},"3633c31e6832",[2384],"f06242a5f1f7","leaderboard",{"_key":2387,"_type":134,"marks":2388,"text":2389},"aea443f75bfb",[],", and 🔗 ",{"_key":2391,"_type":134,"marks":2392,"text":2394},"44c5f83a91c5",[2393],"bd67fdf5c7ad","model weights",{"_key":2396,"_type":134,"marks":2397,"text":2398},"e77c19555e00",[]," to learn more.",[2400,2402,2404],{"_key":2375,"_type":622,"href":2401},"https://arxiv.org/abs/2606.03888",{"_key":2384,"_type":622,"href":2403},"https://github.com/kaiko-ai/eva#radiology",{"_key":2393,"_type":622,"href":2405},"https://huggingface.co/kaiko-ai/coralbay",{"_key":2407,"_type":502},"906c8f045b1f",{"_key":2409,"_type":443,"text":2410},"7e8094857822",[2411,2419,2426],{"_key":2412,"_type":130,"children":2413,"markDefs":2418,"style":1538},"08d8e875db33",[2414],{"_key":2415,"_type":134,"marks":2416,"text":2417},"d50b6c35f89e",[],"CoralBay is released under the MIT License and is intended solely for research purposes. This model has not been validated or certified for clinical use and must not be used to inform, guide, or replace clinical decision-making, diagnosis, treatment, or patient management.",[],{"_key":2420,"_type":130,"children":2421,"markDefs":2425,"style":1538},"28407ab80679",[2422],{"_key":2415,"_type":134,"marks":2423,"text":2424},[],"\nCoralBay's capabilities — including tumor segmentation, volume estimation, lesion tracking, and scan-level diagnosis — have been evaluated in research settings only. Performance may not generalise across patient populations, scanner hardware, imaging protocols, or clinical environments. The model has not been reviewed or approved by any regulatory authority (including but not limited to the FDA or under the EU MDR).",[],{"_key":2427,"_type":130,"children":2428,"markDefs":2432,"style":1538},"85d8875c981d",[2429],{"_key":2415,"_type":134,"marks":2430,"text":2431},[],"\nUsers are solely responsible for assessing suitability for their intended use. Kaiko strongly recommends independent validation, ethical review, and compliance with applicable laws before any downstream application — particularly in healthcare AI contexts. CoralBay is provided \"as is\", without warranty of any kind. Kaiko accepts no liability for outcomes arising from its use.",[],"2026-06-10T09:00:00.000Z",{"_type":89,"asset":2435},{"_ref":2436,"_type":92},"image-0327c08822c6c84776553e305bd13275fb6dc913-2011x941-png",[2438],{"_key":2439,"_ref":2440,"_type":92},"63b393e2b21e","e3bae32e-f001-4ba4-8f4f-e9e38af8cc33",{"_type":1087,"current":2442},"CoralBay-3D-Foundation-Model","Introducing CoralBay: A fully open-source, native 3D foundation model for radiology that delivers state-of-the-art CT scan analysis with unmatched data efficiency.","CoralBay: The 3D Foundation Model for Radiology CT-Scans by kaiko",[2446,2448,2450],{"_key":2447,"_ref":1557,"_type":92},"ab8ee8167c09",{"_key":2449,"_ref":1093,"_type":92},"8e1606b99a64",{"_key":2451,"_ref":2452,"_type":92},"bdc5cc8677ce","fd4ff9cb-2a4f-41a0-a2af-371ea8baafe4",{"documents":2454,"paths":2459,"mappings":2471},[2455,2456,2457,2458],{"_id":1085,"_type":428},{"_id":423,"_type":428},{"_id":1549,"_type":428},{"_id":1105,"_type":428},[145,146,147,148,149,150,2460,2461,2462,2463,2464,2465,2466,2467,2468,2469,2470],"$['area']","$['authors']","$['categories']","$['content']","$['date']","$['featuredImage']","$['relatedPosts']","$['slug']","$['subHeading']","$['title']","$['topics']",{"$[0]['_createdAt']":2472,"$[0]['_id']":2474,"$[0]['_rev']":2476,"$[0]['_system']":2478,"$[0]['_type']":2480,"$[0]['_updatedAt']":2482,"$[0]['area']":2484,"$[0]['authors']":2486,"$[0]['categories']":2488,"$[0]['content']":2490,"$[0]['date']":2492,"$[0]['featuredImage']":2494,"$[0]['relatedPosts']":2496,"$[0]['slug']":2498,"$[0]['subHeading']":2500,"$[0]['title']":2502,"$[0]['topics']":2504,"$[1]['_createdAt']":2506,"$[1]['_id']":2508,"$[1]['_rev']":2510,"$[1]['_system']":2512,"$[1]['_type']":2514,"$[1]['_updatedAt']":2516,"$[1]['area']":2518,"$[1]['authors']":2520,"$[1]['categories']":2522,"$[1]['content']":2524,"$[1]['date']":2526,"$[1]['featuredImage']":2528,"$[1]['relatedPosts']":2530,"$[1]['slug']":2532,"$[1]['subHeading']":2534,"$[1]['title']":2536,"$[1]['topics']":2538,"$[2]['_createdAt']":2540,"$[2]['_id']":2542,"$[2]['_rev']":2544,"$[2]['_system']":2546,"$[2]['_type']":2548,"$[2]['_updatedAt']":2550,"$[2]['area']":2552,"$[2]['categories']":2554,"$[2]['content']":2556,"$[2]['date']":2558,"$[2]['featuredImage']":2560,"$[2]['slug']":2562,"$[2]['title']":2564,"$[2]['topics']":2566,"$[3]['_createdAt']":2568,"$[3]['_id']":2570,"$[3]['_rev']":2572,"$[3]['_system']":2574,"$[3]['_type']":2576,"$[3]['_updatedAt']":2578,"$[3]['area']":2580,"$[3]['authors']":2582,"$[3]['categories']":2584,"$[3]['content']":2586,"$[3]['date']":2588,"$[3]['featuredImage']":2590,"$[3]['relatedPosts']":2592,"$[3]['slug']":2594,"$[3]['subHeading']":2596,"$[3]['title']":2598,"$[3]['topics']":2600},{"source":2473,"type":168},{"document":101,"path":166,"type":167},{"source":2475,"type":168},{"document":101,"path":101,"type":167},{"source":2477,"type":168},{"document":101,"path":173,"type":167},{"source":2479,"type":168},{"document":101,"path":176,"type":167},{"source":2481,"type":168},{"document":101,"path":179,"type":167},{"source":2483,"type":168},{"document":101,"path":182,"type":167},{"source":2485,"type":168},{"document":101,"path":185,"type":167},{"source":2487,"type":168},{"document":101,"path":188,"type":167},{"source":2489,"type":168},{"document":101,"path":191,"type":167},{"source":2491,"type":168},{"document":101,"path":194,"type":167},{"source":2493,"type":168},{"document":101,"path":197,"type":167},{"source":2495,"type":168},{"document":101,"path":200,"type":167},{"source":2497,"type":168},{"document":101,"path":203,"type":167},{"source":2499,"type":168},{"document":101,"path":206,"type":167},{"source":2501,"type":168},{"document":101,"path":209,"type":167},{"source":2503,"type":168},{"document":101,"path":212,"type":167},{"source":2505,"type":168},{"document":101,"path":218,"type":167},{"source":2507,"type":168},{"document":176,"path":166,"type":167},{"source":2509,"type":168},{"document":176,"path":101,"type":167},{"source":2511,"type":168},{"document":176,"path":173,"type":167},{"source":2513,"type":168},{"document":176,"path":176,"type":167},{"source":2515,"type":168},{"document":176,"path":179,"type":167},{"source":2517,"type":168},{"document":176,"path":182,"type":167},{"source":2519,"type":168},{"document":176,"path":185,"type":167},{"source":2521,"type":168},{"document":176,"path":188,"type":167},{"source":2523,"type":168},{"document":176,"path":191,"type":167},{"source":2525,"type":168},{"document":176,"path":194,"type":167},{"source":2527,"type":168},{"document":176,"path":197,"type":167},{"source":2529,"type":168},{"document":176,"path":200,"type":167},{"source":2531,"type":168},{"document":176,"path":203,"type":167},{"source":2533,"type":168},{"document":176,"path":206,"type":167},{"source":2535,"type":168},{"document":176,"path":209,"type":167},{"source":2537,"type":168},{"document":176,"path":212,"type":167},{"source":2539,"type":168},{"document":176,"path":218,"type":167},{"source":2541,"type":168},{"document":173,"path":166,"type":167},{"source":2543,"type":168},{"document":173,"path":101,"type":167},{"source":2545,"type":168},{"document":173,"path":173,"type":167},{"source":2547,"type":168},{"document":173,"path":176,"type":167},{"source":2549,"type":168},{"document":173,"path":179,"type":167},{"source":2551,"type":168},{"document":173,"path":182,"type":167},{"source":2553,"type":168},{"document":173,"path":185,"type":167},{"source":2555,"type":168},{"document":173,"path":191,"type":167},{"source":2557,"type":168},{"document":173,"path":194,"type":167},{"source":2559,"type":168},{"document":173,"path":197,"type":167},{"source":2561,"type":168},{"document":173,"path":200,"type":167},{"source":2563,"type":168},{"document":173,"path":206,"type":167},{"source":2565,"type":168},{"document":173,"path":212,"type":167},{"source":2567,"type":168},{"document":173,"path":218,"type":167},{"source":2569,"type":168},{"document":166,"path":166,"type":167},{"source":2571,"type":168},{"document":166,"path":101,"type":167},{"source":2573,"type":168},{"document":166,"path":173,"type":167},{"source":2575,"type":168},{"document":166,"path":176,"type":167},{"source":2577,"type":168},{"document":166,"path":179,"type":167},{"source":2579,"type":168},{"document":166,"path":182,"type":167},{"source":2581,"type":168},{"document":166,"path":185,"type":167},{"source":2583,"type":168},{"document":166,"path":188,"type":167},{"source":2585,"type":168},{"document":166,"path":191,"type":167},{"source":2587,"type":168},{"document":166,"path":194,"type":167},{"source":2589,"type":168},{"document":166,"path":197,"type":167},{"source":2591,"type":168},{"document":166,"path":200,"type":167},{"source":2593,"type":168},{"document":166,"path":203,"type":167},{"source":2595,"type":168},{"document":166,"path":206,"type":167},{"source":2597,"type":168},{"document":166,"path":209,"type":167},{"source":2599,"type":168},{"document":166,"path":212,"type":167},{"source":2601,"type":168},{"document":166,"path":218,"type":167},1784134278829]