{"id":2564,"date":"2026-04-23T18:22:10","date_gmt":"2026-04-23T18:22:10","guid":{"rendered":"https:\/\/deepinsightai.io\/?p=2564"},"modified":"2026-04-23T18:22:12","modified_gmt":"2026-04-23T18:22:12","slug":"deepseek-starts-updating-frequently","status":"publish","type":"post","link":"https:\/\/deepinsightai.io\/es\/deepseek-starts-updating-frequently\/","title":{"rendered":"DeepSeek comienza a actualizarse con frecuencia: Tile Kernels y DeepEP V2"},"content":{"rendered":"<p>Hace un momento, DeepSeek's <a href=\"https:\/\/deepinsightai.io\/es\/the-fake-star-economy-on-github\/\">GitHub<\/a> empez\u00f3 a actualizarse con frecuencia. Puso en marcha un nuevo repositorio de c\u00f3digo abierto, <strong>N\u00facleos de azulejos<\/strong>, y, al mismo tiempo, actualiz\u00f3 el <strong>DeepEP<\/strong> repositorio, aportando <strong>DeepEP V2<\/strong> en l\u00ednea. Ha pasado menos de una semana desde que DeepSeek actualiz\u00f3 silenciosamente <strong>Mega ME<\/strong> y <strong>Indexador FP4<\/strong> la \u00faltima vez.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">N\u00facleos de azulejos DeepSeek<\/h2>\n\n\n\n<figure data-spectra-id=\"spectra-mobt4mso-77si3j\" class=\"wp-block-image aligncenter size-full\"><img fetchpriority=\"high\" decoding=\"async\" width=\"889\" height=\"471\" src=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59.png\" alt=\"N\u00facleos de azulejos DeepSeek\" class=\"wp-image-2568\" srcset=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59.png 889w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59-300x159.png 300w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59-768x407.png 768w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-59-18x10.png 18w\" sizes=\"(max-width: 889px) 100vw, 889px\" \/><\/figure>\n\n\n\n<p>Enlace: <code>https:\/\/github.com\/deepseek-ai\/TileKernels<\/code><\/p>\n\n\n\n<p>Seg\u00fan la introducci\u00f3n, <strong>N\u00facleos de azulejos<\/strong> son kernels de GPU optimizados para operaciones LLM, construidos con <strong>TileLang<\/strong>. TileLang es un lenguaje espec\u00edfico para expresar kernels de GPU de alto rendimiento en Python, con caracter\u00edsticas como la f\u00e1cil portabilidad, el desarrollo \u00e1gil y la optimizaci\u00f3n autom\u00e1tica.<\/p>\n\n\n\n<p>El rendimiento de los Tile Kernels es extremadamente alto. Como escribi\u00f3 el propio DeepSeek: \u201cLa mayor\u00eda de los kernels de este proyecto ya est\u00e1n cerca del l\u00edmite de rendimiento del hardware en t\u00e9rminos de intensidad de c\u00e1lculo y ancho de banda de memoria. Algunos de ellos ya se han utilizado internamente en escenarios de entrenamiento e inferencia. Sin embargo, a\u00fan no representan las mejores pr\u00e1cticas, y seguimos mejorando la calidad del c\u00f3digo y la documentaci\u00f3n.\u201d<\/p>\n\n\n\n<p>No hay mucha informaci\u00f3n introductoria en el repositorio, pero entre l\u00edneas ya \u201cdesvela\u201d el camino de innovaci\u00f3n arquitect\u00f3nica subyacente de los modelos de pr\u00f3xima generaci\u00f3n de DeepSeek, se\u00f1alando un salto comparable al reciente <a target=\"_blank\" rel=\"noreferrer noopener\" href=\"https:\/\/deepinsightai.io\/es\/hy3-preview-launch\/\">Preestreno de Hy3<\/a>.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Caracter\u00edsticas de DeepSeek Tile Kernels<\/h3>\n\n\n\n<p>\u00c9stas son algunas de las caracter\u00edsticas espec\u00edficas de los Tile Kernels:<\/p>\n\n\n\n<p><strong>Mecanismo de cierre:<\/strong> Selecci\u00f3n y puntuaci\u00f3n de expertos Top-k para el enrutamiento MoE<\/p>\n\n\n\n<p><strong>Enrutamiento MoE:<\/strong> Asignaci\u00f3n de tokens a expertos, expansi\u00f3n\/reducci\u00f3n fusionada y normalizaci\u00f3n de pesos<\/p>\n\n\n\n<p><strong>Cuantizaci\u00f3n:<\/strong> Admite la conversi\u00f3n FP8\/FP4\/E5M6 en los modos por token, por bloque y por canal, y fusiona las operaciones SwiGLU + cuantificaci\u00f3n.<\/p>\n\n\n\n<p><strong>Transponer:<\/strong> Operaciones de transposici\u00f3n por lotes<\/p>\n\n\n\n<p><strong>Engram:<\/strong> N\u00facleos de engramaci\u00f3n, fusi\u00f3n de RMSNorm, propagaci\u00f3n hacia delante\/atr\u00e1s y reducci\u00f3n del gradiente de peso.<\/p>\n\n\n\n<p><strong>Hiperconexi\u00f3n de colectores:<\/strong> N\u00facleos de hiperconexi\u00f3n, incluida la normalizaci\u00f3n Sinkhorn y la divisi\u00f3n\/aplicaci\u00f3n para la mezcla<\/p>\n\n\n\n<p><strong>Modelado:<\/strong> Alto nivel <code>antorcha.autograd.Function<\/code> envoltorios que combinan los n\u00facleos subyacentes en capas entrenables (engram gate, mHC pipeline)<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">DeepSeek EPv2: EP m\u00e1s r\u00e1pido con soporte de Engram, PP y CP<\/h2>\n\n\n\n<figure data-spectra-id=\"spectra-mobt54cl-th2xp4\" class=\"wp-block-image aligncenter size-large\"><img decoding=\"async\" width=\"1024\" height=\"656\" data-src=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-1024x656.png\" alt=\"DeepSeek EPv2: EP m\u00e1s r\u00e1pido con soporte de Engram, PP y CP\" class=\"wp-image-2569 lazyload\" data-srcset=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-1024x656.png 1024w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-300x192.png 300w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-768x492.png 768w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60-18x12.png 18w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-60.png 1069w\" data-sizes=\"(max-width: 1024px) 100vw, 1024px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 1024px; --smush-placeholder-aspect-ratio: 1024\/656;\" \/><\/figure>\n\n\n\n<p>Enlace EPv2: <code>https:\/\/github.com\/deepseek-ai\/DeepEP\/pull\/605<\/code><\/p>\n\n\n\n<p>Hoy mismo, DeepSeek tambi\u00e9n ha publicado la \u00faltima versi\u00f3n de <strong>EPv2<\/strong>, entregando m\u00e1s r\u00e1pido <strong>paralelismo experto (PE)<\/strong> y apoyo a <strong>Engrama \/ paralelismo de canalizaci\u00f3n (PP) \/ paralelismo de contexto (CP)<\/strong>.<\/p>\n\n\n\n<p>A medida que el hardware, las redes y las arquitecturas de los modelos han ido evolucionando junto con los r\u00e1pidos lanzamientos de la industria como <a href=\"https:\/\/deepinsightai.io\/es\/qwen-3-6\/\" target=\"_blank\" rel=\"noreferrer noopener\">Qwen 3.6<\/a>, El anterior DeepEP V1 de DeepSeek ya hab\u00eda acumulado demasiado bagaje hist\u00f3rico y demasiados problemas de rendimiento.<\/p>\n\n\n\n<p>Esta actualizaci\u00f3n reestructura por completo <strong>Paralelismo experto<\/strong>. En comparaci\u00f3n con V1, s\u00f3lo necesita una fracci\u00f3n de los recursos de SM para alcanzar un rendimiento extremo, al tiempo que admite una mayor escala. <strong>ampliaci\u00f3n<\/strong> (dentro de una misma m\u00e1quina) y <strong>ampliaci\u00f3n<\/strong> (en todas las m\u00e1quinas).<\/p>\n\n\n\n<p>Adem\u00e1s, DeepSeek introdujo un sistema experimental <strong>0 SM<\/strong> en esta actualizaci\u00f3n, incluyendo <strong>0 SM Engram<\/strong>, <strong>0 Paralelismo de canalizaci\u00f3n SM (PP)<\/strong>, y <strong>0 Paralelismo contextual SM (CP)<\/strong> Todos los operadores de recogida. Al mismo tiempo, se ha cambiado el backend de <strong>NVSHMEM<\/strong> al encendedor <strong>NCCL Gin<\/strong> backend.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Nuevas funciones de DeepSeek DeepEP V2<\/h3>\n\n\n\n<p>Estas son algunas de las novedades de DeepEP V2:<\/p>\n\n\n\n<p><strong>Totalmente JIT<\/strong><\/p>\n\n\n\n<p><strong>NCCL Gin backend:<\/strong><\/p>\n\n\n\n<p>S\u00f3lo cabecera, extremadamente ligera<\/p>\n\n\n\n<p>Posibilidad de reutilizar los comunicadores NCCL existentes<\/p>\n\n\n\n<p><strong>EPv2:<\/strong><\/p>\n\n\n\n<p>Unifica las API de alto rendimiento y baja latencia en una \u00fanica interfaz y adopta un nuevo dise\u00f1o GEMM.<\/p>\n\n\n\n<p>Admite dominios de escalado mayores, hasta <strong>EP2048<\/strong><\/p>\n\n\n\n<p>Introduce el c\u00e1lculo anal\u00edtico del recuento de SM y QP, por lo que ya no es necesario el autoajuste.<\/p>\n\n\n\n<p>Sigue apoyando tanto <strong>H\u00edbrido<\/strong> y <strong>Directo<\/strong> modo<\/p>\n\n\n\n<p>En las tareas de formaci\u00f3n m\u00e1s antiguas similares a V3, el uso de SM desciende de <strong>24<\/strong> a <strong>4-6<\/strong>, manteniendo el mismo rendimiento o incluso mejor<\/p>\n\n\n\n<p><strong>0 SM Engram<\/strong> (con RDMA)<\/p>\n\n\n\n<p><strong>0 SM PP<\/strong> (con RDMA)<\/p>\n\n\n\n<p><strong>0 SM CP<\/strong> (con Copy Engine)<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Rendimiento de DeepSeek DeepEP V2<\/h2>\n\n\n\n<p>Tras la configuraci\u00f3n de <strong>DeepSeek-V3<\/strong>, Las pruebas se realizaron con la nueva versi\u00f3n y con los par\u00e1metros de <strong>8.000 fichas por lote<\/strong>, <strong>7168 dimensi\u00f3n oculta<\/strong>, <strong>Los 8 mejores expertos<\/strong>, <strong>Env\u00edo FP8<\/strong>, y <strong>Combinaci\u00f3n BF16<\/strong>. Los resultados son los siguientes:<\/p>\n\n\n\n<figure data-spectra-id=\"spectra-mobt5q25-blwgn9\" class=\"wp-block-image aligncenter size-full\"><img decoding=\"async\" width=\"650\" height=\"289\" data-src=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61.png\" alt=\"Rendimiento de DeepSeek DeepEP V2\" class=\"wp-image-2570 lazyload\" data-srcset=\"https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61.png 650w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61-300x133.png 300w, https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/image-61-18x8.png 18w\" data-sizes=\"(max-width: 650px) 100vw, 650px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 650px; --smush-placeholder-aspect-ratio: 650\/289;\" \/><\/figure>\n\n\n\n<p>Nota: Los resultados mostrados son anchos de banda l\u00f3gicos. Por ejemplo, en el caso de <strong>EP 8 x 2<\/strong>, El <strong>90 GB\/s<\/strong> El ancho de banda incluye en realidad el tr\u00e1fico entre GPU locales (rangos locales).<\/p>\n\n\n\n<p>En comparaci\u00f3n con V1, V2 logra hasta <strong>1,3 veces el rendimiento m\u00e1ximo<\/strong>, ahorrando tanto como <strong>4\u00d7 uso de recursos SM <\/strong>-una optimizaci\u00f3n crucial para seguir siendo competitivo en un panorama dominado por pesos pesados como <a href=\"https:\/\/deepinsightai.io\/es\/claude-opus-4-7\/\" target=\"_blank\" rel=\"noreferrer noopener\">Claude Opus 4.7<\/a>.<\/p>\n\n\n\n<p>Por \u00faltimo, un consejo para DeepSeek: date prisa en publicar <strong>V4<\/strong> ya. Todo el mundo se est\u00e1 impacientando.<\/p>\n\n\n\n<p><\/p>","protected":false},"excerpt":{"rendered":"<p>Just now, DeepSeek\u2019s GitHub started updating frequently. It launched and open-sourced a new repository, Tile Kernels, and at the same time updated the DeepEP repository, bringing DeepEP V2 online. It has been less than a week since DeepSeek quietly updated Mega MoE and FP4 Indexer last time. DeepSeek Tile Kernels Link: https:\/\/github.com\/deepseek-ai\/TileKernels According to the [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":2567,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_seopress_robots_primary_cat":"none","_seopress_titles_title":"%%post_title%%","_seopress_titles_desc":"DeepSeek releases Tile Kernels and DeepEP V2 with faster expert parallelism, 1.3\u00d7 performance boost, and major GPU efficiency gains.","_seopress_robots_index":"","_uag_custom_page_level_css":"","site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[2,10],"tags":[],"class_list":["post-2564","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news","category-llm"],"uagb_featured_image_src":{"full":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"thumbnail":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-150x150.png",150,150,true],"medium":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-300x198.png",300,198,true],"medium_large":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-768x508.png",768,508,true],"large":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"1536x1536":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"2048x2048":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek.png",786,520,false],"trp-custom-language-flag":["https:\/\/deepinsightai.io\/wp-content\/uploads\/2026\/04\/deepseek-18x12.png",18,12,true]},"uagb_author_info":{"display_name":"Claude Carter","author_link":"https:\/\/deepinsightai.io\/es\/author\/cloud-han03gmail-com\/"},"uagb_comment_info":0,"uagb_excerpt":"Just now, DeepSeek\u2019s GitHub started updating frequently. It launched and open-sourced a new repository, Tile Kernels, and at the same time updated the DeepEP repository, bringing DeepEP V2 online. It has been less than a week since DeepSeek quietly updated Mega MoE and FP4 Indexer last time. DeepSeek Tile Kernels Link: https:\/\/github.com\/deepseek-ai\/TileKernels According to the&hellip;","_links":{"self":[{"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/posts\/2564","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/comments?post=2564"}],"version-history":[{"count":1,"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/posts\/2564\/revisions"}],"predecessor-version":[{"id":2571,"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/posts\/2564\/revisions\/2571"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/media\/2567"}],"wp:attachment":[{"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/media?parent=2564"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/categories?post=2564"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/deepinsightai.io\/es\/wp-json\/wp\/v2\/tags?post=2564"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}