{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T06:40:50.084Z","headline":"小红书开源 BigMac：多模态大模型训练新范式","description":"小红书技术团队开源 BigMac，一种针对多模态大模型训练的依赖安全嵌套流水线新范式。它以 LLM 流水线为主干，在不打乱执行顺序的前提下嵌入编码器和生成器计算，相比基线实现 1.08x-1.9x 加速，同时保持激活显存有界。BigMac 已作为 dots 多模态模型训练的核心组件投入生产。","url":"https://www.aioga.com/news/cmrvx8ycy0240bipz0pkrvryq/","mainEntityOfPage":"https://www.aioga.com/news/cmrvx8ycy0240bipz0pkrvryq/","datePublished":"2026-07-22T10:04:51.000Z","dateModified":"2026-07-22T10:04:51.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://mp.weixin.qq.com/s/tNJARtn1jIayL5URg87Row","https://aihot.virxact.com/items/cmrvx8ycy0240bipz0pkrvryq"],"canonicalUrl":"https://www.aioga.com/news/cmrvx8ycy0240bipz0pkrvryq/","directAnswer":{"@type":"Answer","text":"小红书技术团队开源 BigMac，将其定位为多模态大模型训练的依赖安全嵌套流水线范式。该方案以 LLM 流水线为主干，嵌入编码器和生成器计算，并已用于 dots 多模态模型训练。","url":"https://www.aioga.com/news/cmrvx8ycy0240bipz0pkrvryq/","dateCreated":"2026-07-22T10:04:51.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"公众号：小红书技术（dots.llm） source article","url":"https://mp.weixin.qq.com/s/tNJARtn1jIayL5URg87Row","datePublished":"2026-07-22T10:04:51.000Z","provider":{"@type":"Organization","name":"公众号：小红书技术（dots.llm）","url":"https://mp.weixin.qq.com/s/tNJARtn1jIayL5URg87Row"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmrvx8ycy0240bipz0pkrvryq","datePublished":"2026-07-22T10:04:51.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmrvx8ycy0240bipz0pkrvryq"}}],"aggregationSource":"公众号：小红书技术（dots.llm）","originalPublisher":{"name":"公众号：小红书技术（dots.llm）","url":"https://mp.weixin.qq.com/s/tNJARtn1jIayL5URg87Row"},"article":{"id":"cmrvx8ycy0240bipz0pkrvryq","slug":"cmrvx8ycy0240bipz0pkrvryq","url":"https://www.aioga.com/news/cmrvx8ycy0240bipz0pkrvryq/","title":"小红书开源 BigMac：多模态大模型训练新范式","title_en":"小红书 dots infra 开源BigMac ： 突破多模态大模型训练的帕累托前沿","summary":"小红书技术团队开源 BigMac，一种针对多模态大模型训练的依赖安全嵌套流水线新范式。它以 LLM 流水线为主干，在不打乱执行顺序的前提下嵌入编码器和生成器计算，相比基线实现 1.08x-1.9x 加速，同时保持激活显存有界。BigMac 已作为 dots 多模态模型训练的核心组件投入生产。","source":"公众号：小红书技术（dots.llm）","sourceUrl":"https://mp.weixin.qq.com/s/tNJARtn1jIayL5URg87Row","aiHotUrl":"https://aihot.virxact.com/items/cmrvx8ycy0240bipz0pkrvryq","publishedAt":"2026-07-22T10:04:51.000Z","category":"产品更新","score":62,"selected":true,"articleBody":["小红书技术团队开源 BigMac，一种针对多模态大模型训练的依赖安全嵌套流水线新范式。","它以 LLM 流水线为主干，在不打乱执行顺序的前提下嵌入编码器和生成器计算，相比基线实现 1.08x-1.9x 加速，同时保持激活显存有界。","BigMac 已作为 dots 多模态模型训练的核心组件投入生产。"],"articleImages":[],"mediaStatus":"none","articleBodyZh":[],"translationStatus":"","bodyOrigin":"summary-fallback","editorial":{"summary":"小红书技术团队开源 BigMac，将其定位为多模态大模型训练的依赖安全嵌套流水线范式。该方案以 LLM 流水线为主干，嵌入编码器和生成器计算，并已用于 dots 多模态模型训练。","background":"公开材料显示，BigMac 面向多模态大模型训练中的流水线组织问题，在不打乱执行顺序的前提下嵌入编码器和生成器计算，同时保持激活显存有界。材料未披露具体硬件、基线配置及测试条件。","viewpoint":"Aioga 判断，BigMac 的主要看点不是单一加速数字，而是以依赖安全方式组织多类计算，并进入实际生产训练流程。其公开价值可能在于为多模态训练流水线提供一种可复用的实现思路。","implications":"值得关注的是，公开材料给出的相对基线加速范围为 1.08x-1.9x，但缺少不同模型、硬件和任务下的细分结果。Aioga 判断，外部团队能否获得相近收益，仍可能取决于具体训练配置。","nextStep":"后续值得关注开源代码、配置和复现实验是否完整披露，以及不同模型规模、编码器与生成器组合下的表现。Aioga 判断，对该方案的进一步评价应以可复现基线、显存数据和适用边界为依据。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-07-23T03:31:28.236Z","sourceHash":"2328d38573075489","review":{"approved":true,"groundedness":96,"clarity":92,"duplicationRisk":24,"blockingIssues":[],"notes":["候选内容准确覆盖了 BigMac 的开源、技术范式、流水线组织方式、加速范围、激活显存有界及投入生产等来源信息。","关于公开价值、外部团队收益和后续评价标准的内容均明确标注为“Aioga 判断”或关注方向，没有将观点冒充已证实事实。","summary 与 background 对“嵌入编码器和生成器计算”略有重复，但不构成阻断问题。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["产品更新","公众号：小红书技术（dots.llm）"],"translations":{"zh-CN":{"title":"小红书开源 BigMac：多模态大模型训练新范式","summary":"小红书技术团队开源 BigMac，一种针对多模态大模型训练的依赖安全嵌套流水线新范式。它以 LLM 流水线为主干，在不打乱执行顺序的前提下嵌入编码器和生成器计算，相比基线实现 1.08x-1.9x 加速，同时保持激活显存有界。BigMac 已作为 dots 多模态模型训练的核心组件投入生产。","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"小红书开源 BigMac：多模态大模型训练新范式 - Aioga AI资讯","description":"小红书技术团队开源 BigMac，一种针对多模态大模型训练的依赖安全嵌套流水线新范式。它以 LLM 流水线为主干，在不打乱执行顺序的前提下嵌入编码器和生成器计算，相比基线实现 1.08x-1.9x 加速，同时保持激活显存有界。BigMac 已作为 dots 多模态模型训练的核心组件投入生产。","url":"https://www.aioga.com/news/cmrvx8ycy0240bipz0pkrvryq/"},"en":{"title":"Xiaohongshu Open Source BigMac: A New Paradigm for Multimodal Large Model Training","summary":"Xiaohongshu's technology team has open-sourced BigMac, a new paradigm for secure dependency nested pipelines in multi-modal large model training. It uses an LLM pipeline as the backbone, embedding encoder and generator computations without disrupting the execution order. Compared to the baseline, it achieves 1.08x-1.9x speedup while keeping the activation GPU memory bounded. BigMac has already been deployed as a core component in the production training of the dots multi-modal model.","category":"Products","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu Open Source BigMac: A New Paradigm for Multimodal Large Model Training - Aioga AI News","description":"Xiaohongshu's technology team has open-sourced BigMac, a new paradigm for secure dependency nested pipelines in multi-modal large model training. It uses an LLM pipeline as the bac...","url":"https://www.aioga.com/en/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:25:14.143Z"},"ja":{"title":"小紅書オープンソース BigMac：マルチモーダル大規模モデル訓練の新しいパラダイム","summary":"小紅書の技術チームは BigMac をオープンソース化しました。これはマルチモーダル大規模モデルのトレーニング向けの依存安全なネスト型パイプラインの新しいパラダイムです。LLM パイプラインを主幹とし、実行順序を乱さずにエンコーダーとジェネレーターの計算を組み込むことで、基準実装に比べて 1.08x-1.9x の高速化を実現しつつ、アクティベーションメモリを制御しています。BigMac はすでに dots マルチモーダルモデルのトレーニングのコアコンポーネントとして本番投入されています。","category":"製品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"小紅書オープンソース BigMac：マルチモーダル大規模モデル訓練の新しいパラダイム - Aioga AIニュース","description":"小紅書の技術チームは BigMac をオープンソース化しました。これはマルチモーダル大規模モデルのトレーニング向けの依存安全なネスト型パイプラインの新しいパラダイムです。LLM パイプラインを主幹とし、実行順序を乱さずにエンコーダーとジェネレーターの計算を組み込むことで、基準実装に比べて 1.08x-1.9x の高速化を実現しつつ、アクティベーションメモリを...","url":"https://www.aioga.com/ja/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:25:34.176Z"},"ko":{"title":"샤오홍슈 오픈소스 BigMac: 다중 모달 대형 모델 훈련의 새로운 패러다임","summary":"샤오홍슈 기술 팀이 BigMac을 오픈소스로 공개했습니다. 이는 다중 모달 대형 모델 훈련을 위한 의존성 안전 중첩 파이프라인의 새로운 패러다임입니다. LLM 파이프라인을 주축으로 하여 실행 순서를 방해하지 않고 인코더와 생성기 계산을 삽입하며, 기준 대비 1.08배에서 1.9배까지 가속을 구현하면서 활성화 메모리 사용을 제한합니다. BigMac은 이미 dots 다중 모달 모델 훈련의 핵심 구성 요소로 생산 환경에 투입되었습니다.","category":"제품 업데이트","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"샤오홍슈 오픈소스 BigMac: 다중 모달 대형 모델 훈련의 새로운 패러다임 - Aioga AI 뉴스","description":"샤오홍슈 기술 팀이 BigMac을 오픈소스로 공개했습니다. 이는 다중 모달 대형 모델 훈련을 위한 의존성 안전 중첩 파이프라인의 새로운 패러다임입니다. LLM 파이프라인을 주축으로 하여 실행 순서를 방해하지 않고 인코더와 생성기 계산을 삽입하며, 기준 대비 1.08배에서 1.9배까지 가속을 구현하면서 활성화 메모리 사용을...","url":"https://www.aioga.com/ko/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:26:18.917Z"},"es":{"title":"Xiaohongshu lanza BigMac de código abierto: nuevo paradigma de entrenamiento de modelos multimodales","summary":"El equipo técnico de Xiaohongshu ha abierto BigMac, un nuevo paradigma de canalización anidada de seguridad de dependencias para el entrenamiento de grandes modelos multimodales. Utiliza la canalización LLM como columna vertebral, incorporando el cálculo del codificador y del generador sin alterar el orden de ejecución, logrando una aceleración de 1,08x-1,9x en comparación con la línea base, al mismo tiempo que mantiene la memoria de activación limitada. BigMac ya se ha implementado en producción como el componente central del entrenamiento del modelo multimodal dots.","category":"Productos","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu lanza BigMac de código abierto: nuevo paradigma de entrenamiento de modelos multimodales - Aioga Noticias de IA","description":"El equipo técnico de Xiaohongshu ha abierto BigMac, un nuevo paradigma de canalización anidada de seguridad de dependencias para el entrenamiento de grandes modelos multimodales. U...","url":"https://www.aioga.com/es/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:26:15.638Z"},"fr":{"title":"Xiaohongshu open source BigMac : nouveau paradigme de formation de modèles multimodaux","summary":"L'équipe technique de Xiaohongshu a rendu open source BigMac, un nouveau modèle de pipeline imbriqué sécurisé pour les dépendances, destiné à l'entraînement de grands modèles multimodaux. Il utilise le pipeline LLM comme tronc principal et intègre les calculs des encodeurs et des générateurs sans perturber l'ordre d'exécution, offrant une accélération de 1,08 à 1,9 fois par rapport à la base, tout en maintenant une mémoire d'activation limitée. BigMac a déjà été déployé en production en tant que composant central de l'entraînement du modèle multimodal dots.","category":"Produits","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu open source BigMac : nouveau paradigme de formation de modèles multimodaux - Aioga Actualités IA","description":"L'équipe technique de Xiaohongshu a rendu open source BigMac, un nouveau modèle de pipeline imbriqué sécurisé pour les dépendances, destiné à l'entraînement de grands modèles multi...","url":"https://www.aioga.com/fr/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:27:08.479Z"},"de":{"title":"Xiaohongshu Open Source BigMac: Neues Paradigma für multimodales Large-Model-Training","summary":"Das technische Team von Xiaohongshu hat BigMac als Open-Source veröffentlicht, ein neues Paradigma für sichere verschachtelte Pipelines beim Training multimodaler großer Modelle. Es verwendet die LLM-Pipeline als Hauptstruktur und bettet Encoder- und Generatorberechnungen ein, ohne die Ausführungsreihenfolge zu stören. Im Vergleich zur Basisimplementierung erreicht es eine Beschleunigung von 1,08x bis 1,9x und hält gleichzeitig den Aktivspeicher begrenzt. BigMac wurde bereits als Kernkomponente für das Training des multimodalen Modells dots in der Produktion eingesetzt.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu Open Source BigMac: Neues Paradigma für multimodales Large-Model-Training - Aioga KI-News","description":"Das technische Team von Xiaohongshu hat BigMac als Open-Source veröffentlicht, ein neues Paradigma für sichere verschachtelte Pipelines beim Training multimodaler großer Modelle. E...","url":"https://www.aioga.com/de/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:27:00.487Z"},"pt-BR":{"title":"Xiaohongshu Open Source BigMac: Novo Paradigma de Treinamento de Modelos Grandes Multimodais","summary":"A equipe de tecnologia do Xiaohongshu tornou o BigMac de código aberto, um novo paradigma de pipeline aninhada com segurança de dependência para treinamento de grandes modelos multimodais. Ele utiliza o pipeline LLM como espinha dorsal, incorporando cálculos de codificador e gerador sem perturbar a ordem de execução, oferecendo uma aceleração de 1,08x a 1,9x em comparação com a baseline, ao mesmo tempo em que mantém a memória ativa dentro de limites. O BigMac já foi implementado como componente central no treinamento do modelo multimodal dots.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu Open Source BigMac: Novo Paradigma de Treinamento de Modelos Grandes Multimodais - Aioga Notícias de IA","description":"A equipe de tecnologia do Xiaohongshu tornou o BigMac de código aberto, um novo paradigma de pipeline aninhada com segurança de dependência para treinamento de grandes modelos mult...","url":"https://www.aioga.com/pt-BR/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:27:50.340Z"},"ru":{"title":"Xiaohongshu открывает исходный код BigMac: новый подход к обучению многомодальных больших моделей","summary":"Техническая команда XiaoHongShu открыла исходный код BigMac, новой парадигмы безопасных вложенных конвейеров для обучения мультимодальных больших моделей. Она использует LLM-конвейер в качестве основного каркаса, внедряя вычисления кодировщика и генератора без нарушения порядка выполнения, что обеспечивает ускорение на 1,08x-1,9x по сравнению с базовой реализацией, при этом оставаясь в пределах допустимого объема видеопамяти для активаций. BigMac уже используется как основной компонент в обучении мультимодальной модели dots в производстве.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu открывает исходный код BigMac: новый подход к обучению многомодальных больших моделей - Aioga Новости ИИ","description":"Техническая команда XiaoHongShu открыла исходный код BigMac, новой парадигмы безопасных вложенных конвейеров для обучения мультимодальных больших моделей. Она использует LLM-конвей...","url":"https://www.aioga.com/ru/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:27:56.106Z"},"ar":{"title":"شياو هونغ شو تفتح المصدر BigMac: نموذج تدريب جديد للنماذج الكبيرة متعددة الوسائط","summary":"أصدر فريق التكنولوجيا في XiaoHongShu مشروع BigMac مفتوح المصدر، وهو نموذج جديد لسلسلة معالجة متداخلة آمنة للاعتماديات لتدريب النماذج الكبيرة متعددة الوسائط. يعتمد على سلسلة معالجة LLM كهيكل رئيسي، ويقوم بإدراج حسابات المشفر والمولّد دون تعطيل ترتيب التنفيذ، بالمقارنة مع الأساسيات يحقق تسريعاً بمعدل 1.08x-1.9x، مع الحفاظ على ذاكرة التفعيل محدودة. تم استخدام BigMac بالفعل كعنصر أساسي في تدريب نموذج dots متعدد الوسائط في الإنتاج.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"شياو هونغ شو تفتح المصدر BigMac: نموذج تدريب جديد للنماذج الكبيرة متعددة الوسائط - Aioga أخبار الذكاء الاصطناعي","description":"أصدر فريق التكنولوجيا في XiaoHongShu مشروع BigMac مفتوح المصدر، وهو نموذج جديد لسلسلة معالجة متداخلة آمنة للاعتماديات لتدريب النماذج الكبيرة متعددة الوسائط. يعتمد على سلسلة معالجة...","url":"https://www.aioga.com/ar/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:28:43.220Z"},"hi":{"title":"शाओहóngबुक ओपन सोर्स BigMac: बहु-मोडल बड़े मॉडल प्रशिक्षण का नया पैटर्न","summary":"Xiaohongshu तकनीकी टीम ने BigMac को ओपन सोर्स किया है, जो बहु-मोडल बड़े मॉडल प्रशिक्षण के लिए एक निर्भरता-सुरक्षित नेस्टेड पाइपलाइन का नया पैमाना है। यह LLM पाइपलाइन को मुख्य धारा के रूप में उपयोग करता है, और निष्पादन क्रम को बाधित किए बिना एन्कोडर और जेनरेटर की गणना को सम्मिलित करता है। बेसलाइन की तुलना में यह 1.08x-1.9x की गति बढ़ाता है, जबकि सक्रिय मेमोरी सीमित रहती है। BigMac पहले से ही Dots बहु-मोडल मॉडल प्रशिक्षण के मुख्य घटक के रूप में उत्पादन में लागू किया जा चुका है।","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"शाओहóngबुक ओपन सोर्स BigMac: बहु-मोडल बड़े मॉडल प्रशिक्षण का नया पैटर्न - Aioga AI समाचार","description":"Xiaohongshu तकनीकी टीम ने BigMac को ओपन सोर्स किया है, जो बहु-मोडल बड़े मॉडल प्रशिक्षण के लिए एक निर्भरता-सुरक्षित नेस्टेड पाइपलाइन का नया पैमाना है। यह LLM पाइपलाइन को मुख्य धारा...","url":"https://www.aioga.com/hi/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:28:53.531Z"},"it":{"title":"Xiaohongshu apre BigMac: nuovo paradigma di addestramento di grandi modelli multimodali","summary":"Il team tecnico di Xiaohongshu ha reso open source BigMac, un nuovo paradigma di pipeline annidata sicura per le dipendenze, destinato all'addestramento di modelli di grandi dimensioni multimodali. Si basa su una pipeline LLM come struttura principale, integrando calcoli di encoder e generator senza interrompere l'ordine di esecuzione, ottenendo un'accelerazione da 1,08x a 1,9x rispetto alla linea di base, mantenendo al contempo la memoria attiva limitata. BigMac è già stato implementato come componente centrale per l'addestramento del modello multimodale dots in produzione.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu apre BigMac: nuovo paradigma di addestramento di grandi modelli multimodali - Aioga Notizie IA","description":"Il team tecnico di Xiaohongshu ha reso open source BigMac, un nuovo paradigma di pipeline annidata sicura per le dipendenze, destinato all'addestramento di modelli di grandi dimens...","url":"https://www.aioga.com/it/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:29:48.843Z"},"nl":{"title":"Xiaohongshu open source BigMac: een nieuw paradigma voor multimodale grote modeltraining","summary":"Het technische team van Xiaohongshu heeft BigMac open source gemaakt, een nieuw paradigma van afhankelijkheidsveilige geneste pipelines voor het trainen van multimodale grote modellen. Het gebruikt de LLM-pipeline als hoofdstructuur en integreert de berekeningen van encoders en generatoren zonder de uitvoeringsvolgorde te verstoren, waardoor het, in vergelijking met de basislijn, 1,08x-1,9x versnelling bereikt en tegelijkertijd de activaties in het VRAM beperkt houdt. BigMac wordt al gebruikt als kerncomponent voor de training van het multimodale model dots in productie.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu open source BigMac: een nieuw paradigma voor multimodale grote modeltraining - Aioga AI-nieuws","description":"Het technische team van Xiaohongshu heeft BigMac open source gemaakt, een nieuw paradigma van afhankelijkheidsveilige geneste pipelines voor het trainen van multimodale grote model...","url":"https://www.aioga.com/nl/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:29:38.268Z"},"tr":{"title":"Xiaohongshu Açık Kaynak BigMac: Çok Modlu Büyük Model Eğitiminde Yeni Paradigma","summary":"Xiaohongshu teknik ekibi, çok modlu büyük model eğitimi için güvenli bağımlılık iç içe boru hattı yeni bir paradigması olan BigMac'i açık kaynak yaptı. Bu sistem, LLM boru hattını ana hat olarak kullanır ve yürütme sırasını bozmayacak şekilde kodlayıcı ve üretici hesaplamalarını entegre eder; baz hattına kıyasla 1,08x-1,9x hızlanma sağlar ve aynı zamanda etkin bellek kullanımını sınırlı tutar. BigMac, dots çok modlu model eğitiminde temel bileşen olarak üretimde kullanılmaktadır.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu Açık Kaynak BigMac: Çok Modlu Büyük Model Eğitiminde Yeni Paradigma - Aioga AI Haberleri","description":"Xiaohongshu teknik ekibi, çok modlu büyük model eğitimi için güvenli bağımlılık iç içe boru hattı yeni bir paradigması olan BigMac'i açık kaynak yaptı. Bu sistem, LLM boru hattını...","url":"https://www.aioga.com/tr/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:30:44.784Z"},"vi":{"title":"Xiaohongshu Mở mã nguồn BigMac: Mô hình huấn luyện lớn đa phương thức mới","summary":"Đội ngũ kỹ thuật Xiaohongshu đã mở nguồn BigMac, một phương thức mới về dây chuyền liên tục lồng nhau an toàn phụ thuộc, dành cho huấn luyện mô hình lớn đa phương thức. Nó lấy dây chuyền LLM làm xương sống, nhúng tính toán bộ mã hóa và bộ sinh mà không làm xáo trộn thứ tự thực thi, so với baseline đạt tốc độ tăng 1.08x-1.9x, đồng thời duy trì bộ nhớ một cách có giới hạn. BigMac đã được đưa vào sản xuất như thành phần cốt lõi của việc huấn luyện mô hình đa phương thức dots.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu Mở mã nguồn BigMac: Mô hình huấn luyện lớn đa phương thức mới - Tin tức AI Aioga","description":"Đội ngũ kỹ thuật Xiaohongshu đã mở nguồn BigMac, một phương thức mới về dây chuyền liên tục lồng nhau an toàn phụ thuộc, dành cho huấn luyện mô hình lớn đa phương thức. Nó lấy dây...","url":"https://www.aioga.com/vi/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:30:41.466Z"},"id":{"title":"Xiaohongshu Merilis BigMac: Paradigma Baru Pelatihan Model Besar Multimodal","summary":"Tim teknologi XiaoHongShu merilis BigMac secara open source, sebuah paradigma baru untuk pipeline bersarang yang aman dalam ketergantungan untuk pelatihan model besar multimodal. Ini menggunakan pipeline LLM sebagai kerangka utama, menyematkan perhitungan encoder dan generator tanpa mengganggu urutan eksekusi, dengan akselerasi 1,08x-1,9x dibandingkan baseline, sambil menjaga penggunaan memori aktif tetap terbatas. BigMac telah digunakan sebagai komponen inti dalam pelatihan model multimodal dots.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu Merilis BigMac: Paradigma Baru Pelatihan Model Besar Multimodal - Berita AI Aioga","description":"Tim teknologi XiaoHongShu merilis BigMac secara open source, sebuah paradigma baru untuk pipeline bersarang yang aman dalam ketergantungan untuk pelatihan model besar multimodal. I...","url":"https://www.aioga.com/id/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:31:30.131Z"},"th":{"title":"Xiaohongshu เปิดตัว BigMac แบบโอเพ่นซอร์ส: รูปแบบใหม่ในการฝึกโมเดลขนาดใหญ่หลากหลายโหมด","summary":"ทีมเทคนิคของ Xiaohongshu เปิดซอร์ส BigMac ซึ่งเป็นรูปแบบใหม่ของสายการผลิตที่ปลอดภัยต่อการพึ่งพาสำหรับการฝึกโมเดลขนาดใหญ่หลายโหมด โดยมี LLM pipeline เป็นแกนหลัก ในขณะที่ฝังการคำนวณของ encoder และ generator โดยไม่รบกวนลำดับการประมวลผล เมื่อเทียบกับฐาน จะทำให้เร็วขึ้น 1.08x-1.9x พร้อมกับรักษาการใช้หน่วยความจำแอคทีฟให้อยู่ในขอบเขต BigMac ได้ถูกใช้เป็นส่วนประกอบหลักในการฝึกโมเดลหลายโหมดของ dots ในการผลิตจริงแล้ว","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu เปิดตัว BigMac แบบโอเพ่นซอร์ส: รูปแบบใหม่ในการฝึกโมเดลขนาดใหญ่หลากหลายโหมด - ข่าว AI Aioga","description":"ทีมเทคนิคของ Xiaohongshu เปิดซอร์ส BigMac ซึ่งเป็นรูปแบบใหม่ของสายการผลิตที่ปลอดภัยต่อการพึ่งพาสำหรับการฝึกโมเดลขนาดใหญ่หลายโหมด โดยมี LLM pipeline เป็นแกนหลัก ในขณะที่ฝังการคำนวณข...","url":"https://www.aioga.com/th/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:31:37.260Z"},"pl":{"title":"Xiaohongshu otwiera BigMac: nowy paradygmat treningu dużych modeli multimodalnych","summary":"Zespół techniczny Xiaohongshu udostępnił jako open source BigMac, nowy paradygmat bezpiecznego, zależnościowego, zagnieżdżonego potoku dla treningu dużych modeli multimodalnych. Bazuje on na potoku LLM jako głównym szkielecie, a przy niezmienionym porządku wykonywania wprowadza obliczenia enkodera i generatora, co w porównaniu do bazowej implementacji przyspiesza 1,08-1,9 razy, jednocześnie utrzymując ograniczone zużycie pamięci aktywnej. BigMac został wdrożony w produkcji jako kluczowy komponent treningu multimodalnego modelu dots.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu otwiera BigMac: nowy paradygmat treningu dużych modeli multimodalnych - Aioga Wiadomości AI","description":"Zespół techniczny Xiaohongshu udostępnił jako open source BigMac, nowy paradygmat bezpiecznego, zależnościowego, zagnieżdżonego potoku dla treningu dużych modeli multimodalnych. Ba...","url":"https://www.aioga.com/pl/news/cmrvx8ycy0240bipz0pkrvryq/","contentTranslated":true,"sourceHash":"7f62d3b95e001db6","translatedAt":"2026-07-22T17:32:23.427Z"}}}}