{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T06:40:50.084Z","headline":"腾讯混元发布 HyOCR-1.5：端到端 OCR 大模型全栈开源，推理提速 6.37 倍","description":"腾讯混元发布 HyOCR-1.5，这是端到端 OCR 大模型领域首个将训练、推理、模型权重完整开源的专家模型。仅 1B 参数，覆盖 8 种以上 text-centric 任务。引入 DFlash 投机解码框架，在 Transformers 下实现 6.37× 加速，vLLM 下 2.14× 加速，端到端推理达每页 1.408s。支持 4K 分辨率与 128K 上下文窗口，通过 Agentic Data Flow 扩展低资源 OCR（331 种语言）、古文字识别与多图问答能力。在 OmniDocBench v1.6 上以 94.74 分居端到端第一。","url":"https://www.aioga.com/news/cmrj4s89p05nhbilkljqel2d5/","mainEntityOfPage":"https://www.aioga.com/news/cmrj4s89p05nhbilkljqel2d5/","datePublished":"2026-07-13T11:12:59.000Z","dateModified":"2026-07-13T11:12:59.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://mp.weixin.qq.com/s/vKFCa9FfoGBUGK8J1MhFag","https://aihot.virxact.com/items/cmrj4s89p05nhbilkljqel2d5"],"canonicalUrl":"https://www.aioga.com/news/cmrj4s89p05nhbilkljqel2d5/","directAnswer":{"@type":"Answer","text":"腾讯混元发布端到端 OCR 专家模型 HyOCR-1.5，并开放训练、推理与模型权重。模型参数量为 1B，覆盖 8 种以上文本中心任务，在 OmniDocBench v1.6 上以 94.74 分位居端到端第一。","url":"https://www.aioga.com/news/cmrj4s89p05nhbilkljqel2d5/","dateCreated":"2026-07-13T11:12:59.000Z","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：腾讯混元 source article","url":"https://mp.weixin.qq.com/s/vKFCa9FfoGBUGK8J1MhFag","datePublished":"2026-07-13T11:12:59.000Z","provider":{"@type":"Organization","name":"公众号：腾讯混元","url":"https://mp.weixin.qq.com/s/vKFCa9FfoGBUGK8J1MhFag"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmrj4s89p05nhbilkljqel2d5","datePublished":"2026-07-13T11:12:59.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmrj4s89p05nhbilkljqel2d5"}}],"aggregationSource":"公众号：腾讯混元","originalPublisher":{"name":"公众号：腾讯混元","url":"https://mp.weixin.qq.com/s/vKFCa9FfoGBUGK8J1MhFag"},"article":{"id":"cmrj4s89p05nhbilkljqel2d5","slug":"cmrj4s89p05nhbilkljqel2d5","url":"https://www.aioga.com/news/cmrj4s89p05nhbilkljqel2d5/","title":"腾讯混元发布 HyOCR-1.5：端到端 OCR 大模型全栈开源，推理提速 6.37 倍","title_en":"训推全开源：HyOCR-1.5 让端到端 OCR 大模型跑得更快、看得更准、功能更全","summary":"腾讯混元发布 HyOCR-1.5，这是端到端 OCR 大模型领域首个将训练、推理、模型权重完整开源的专家模型。仅 1B 参数，覆盖 8 种以上 text-centric 任务。引入 DFlash 投机解码框架，在 Transformers 下实现 6.37× 加速，vLLM 下 2.14× 加速，端到端推理达每页 1.408s。支持 4K 分辨率与 128K 上下文窗口，通过 Agentic Data Flow 扩展低资源 OCR（331 种语言）、古文字识别与多图问答能力。在 OmniDocBench v1.6 上以 94.74 分居端到端第一。","source":"公众号：腾讯混元","sourceUrl":"https://mp.weixin.qq.com/s/vKFCa9FfoGBUGK8J1MhFag","aiHotUrl":"https://aihot.virxact.com/items/cmrj4s89p05nhbilkljqel2d5","publishedAt":"2026-07-13T11:12:59.000Z","category":"模型更新","score":76,"selected":true,"articleBody":["腾讯混元发布 HyOCR-1.5，这是端到端 OCR 大模型领域首个将训练、推理、模型权重完整开源的专家模型。","仅 1B 参数，覆盖 8 种以上 text-centric 任务。","引入 DFlash 投机解码框架，在 Transformers 下实现 6.37× 加速，vLLM 下 2.14× 加速，端到端推理达每页 1.408s。","支持 4K 分辨率与 128K 上下文窗口，通过 Agentic Data Flow 扩展低资源 OCR（331 种语言）、古文字识别与多图问答能力。","在 OmniDocBench v1.6 上以 94.74 分居端到端第一。"],"articleImages":[],"mediaStatus":"none","articleBodyZh":[],"translationStatus":"","bodyOrigin":"summary-fallback","editorial":{"summary":"腾讯混元发布端到端 OCR 专家模型 HyOCR-1.5，并开放训练、推理与模型权重。模型参数量为 1B，覆盖 8 种以上文本中心任务，在 OmniDocBench v1.6 上以 94.74 分位居端到端第一。","background":"HyOCR-1.5 支持 4K 分辨率和 128K 上下文窗口，并通过 Agentic Data Flow 扩展至低资源 OCR、古文字识别和多图问答。其中低资源 OCR 覆盖 331 种语言。","viewpoint":"Aioga 判断，HyOCR-1.5 的关注点不仅是榜单成绩，还在于训练、推理和模型权重的完整开放。1B 参数与多项推理加速结果，也可能提升其在研究复现及实际部署评估中的吸引力。","implications":"值得关注的是，DFlash 投机解码框架在 Transformers 下实现 6.37 倍加速，在 vLLM 下实现 2.14 倍加速，端到端推理达到每页 1.408 秒，可能为 OCR 应用的性能评估提供新参照。","nextStep":"后续可重点核验开源训练与推理代码的完整度、不同硬件和文档类型下的速度表现，以及 331 种语言、古文字识别和多图问答在公开测试之外的稳定性与可复现性。","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":"2026-07-22T19:51:55.489Z","sourceHash":"a258c8bbbaa622d2","review":{"approved":true,"groundedness":96,"clarity":92,"duplicationRisk":28,"blockingIssues":[],"notes":["候选内容中的参数规模、任务覆盖、基准成绩、分辨率、上下文窗口、语言数量及推理加速数据均与来源材料一致。","“可能提升其在研究复现及实际部署评估中的吸引力”和“可能为 OCR 应用的性能评估提供新参照”属于合理推测，且已通过“Aioga 判断”“可能”等措辞明确标示为观点，并未冒充来源事实。","nextStep 属于后续核验建议，不构成事实断言。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["模型更新","公众号：腾讯混元"],"translations":{"zh-CN":{"title":"腾讯混元发布 HyOCR-1.5：端到端 OCR 大模型全栈开源，推理提速 6.37 倍","summary":"腾讯混元发布 HyOCR-1.5，这是端到端 OCR 大模型领域首个将训练、推理、模型权重完整开源的专家模型。仅 1B 参数，覆盖 8 种以上 text-centric 任务。引入 DFlash 投机解码框架，在 Transformers 下实现 6.37× 加速，vLLM 下 2.14× 加速，端到端推理达每页 1.408s。支持 4K 分辨率与 128K 上下文窗口，通过 Agentic Data Flow 扩展低资源 OCR（331 种语言）、古文字识别与多图问答能力。在 OmniDocBench v1.6 上以 94.74 分居端到端第一。","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"腾讯混元发布 HyOCR-1.5：端到端 OCR 大模型全栈开源，推理提速 6.37 倍 - Aioga AI资讯","description":"腾讯混元发布 HyOCR-1.5，这是端到端 OCR 大模型领域首个将训练、推理、模型权重完整开源的专家模型。仅 1B 参数，覆盖 8 种以上 text-centric 任务。引入 DFlash 投机解码框架，在 Transformers 下实现 6.37× 加速，vLLM 下 2.14× 加速，端到端推理达每页 1.408s。支持 4K 分辨率与 128K","url":"https://www.aioga.com/news/cmrj4s89p05nhbilkljqel2d5/"},"en":{"title":"Tencent Hunyuan released HyOCR-1.5: an end-to-end OCR large model full-stack open-source with 6.37x faster inference","summary":"Tencent Hunyuan released HyOCR-1.5, the first expert model in the end-to-end OCR large model field to fully open source training, inference, and model weighting. With just 1B parameters, it covers more than 8 text-centric tasks. Introduced the DFlash speculative decoding framework, achieving 6.37× acceleration under Transformers and 2.14× acceleration under vLLM, with end-to-end inference reaching 1.408 seconds per page. Supports 4K resolution and 128K context windows, and expands low-resource OCR (331 languages), ancient script recognition, and multi-image Q&A capabilities through Agentic Data Flow. On OmniDocBench v1.6, it ranked first end-to-end with a score of 94.74.","category":"Models","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan released HyOCR-1.5: an end-to-end OCR large model full-stack open-source with 6.37x faster inference - Aioga AI News","description":"Tencent Hunyuan released HyOCR-1.5, the first expert model in the end-to-end OCR large model field to fully open source training, inference, and model weighting. With just 1B param","url":"https://www.aioga.com/en/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:04.641Z"},"ja":{"title":"騰訊環源はHyOCR-1.5をリリースしました。これは、6.37倍の推論速度を持つエンドツーエンドの大規模モデルフルスタックオープンソースです","summary":"騰訊環源はHyOCR-1.5をリリースしました。これは、エンドツーエンドのOCR大規模モデル分野で初めて完全オープンソースのトレーニング、推論、モデル重み付けを備えたエキスパートモデルです。 1Bパラメータだけで、8以上のテキスト中心のタスクをカバーしています。 DFlashの推測的復号フレームワークを導入し、トランスフォーマーで6.37×の加速、vLLMで2.14×の加速を実現し、エンドツーエンドの推論時間は1ページあたり1.408秒に達しました。 4K解像度と128Kコンテキストウィンドウをサポートし、Agentic Data Flowを通じて低リソースのOCR(331言語)、古代スクリプト認識、多画像Q&A機能を拡張します。 OmniDocBench v1.6では、94.74点でエンドツーエンドで1位にランクされました。","category":"モデル更新","source":"公众号：腾讯混元","pageTitle":"騰訊環源はHyOCR-1.5をリリースしました。これは、6.37倍の推論速度を持つエンドツーエンドの大規模モデルフルスタックオープンソースです - Aioga AIニュース","description":"騰訊環源はHyOCR-1.5をリリースしました。これは、エンドツーエンドのOCR大規模モデル分野で初めて完全オープンソースのトレーニング、推論、モデル重み付けを備えたエキスパートモデルです。 1Bパラメータだけで、8以上のテキスト中心のタスクをカバーしています。 DFlashの推測的復号フレームワークを導入し、トランスフォーマーで6.37×の加速、vLLMで","url":"https://www.aioga.com/ja/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:04.798Z"},"ko":{"title":"텐센트 훈위안은 HyOCR-1.5를 출시했습니다: 6.37배 빠른 추론을 제공하는 종단 간 OCR 대형 모델 풀스택 오픈소스","summary":"텐센트 훈위안은 완전 오픈 소스 학습, 추론, 모델 가중치를 갖춘 엔드 투 엔드 OCR 대형 모델 분야에서 최초의 전문가 모델인 HyOCR-1.5를 출시했습니다. 1B 매개변수만으로 8개 이상의 텍스트 중심 작업을 다룹니다. DFlash 추측적 디코딩 프레임워크를 도입하여 트랜스포머에서 6.37× 가속, vLLM에서 2.14× 가속을 달성했으며, 종단 간 추론은 페이지당 1.408초에 달합니다. 4K 해상도와 128K 컨텍스트 창을 지원하며, 에이전트 데이터 플로우를 통해 낮은 자원의 OCR(331개 언어), 고대 스크립트 인식, 다중 이미지 Q&A 기능을 확장합니다. OmniDocBench v1.6에서는 94.74점으로 엔드 투 엔드 1위를 차지했습니다.","category":"모델 업데이트","source":"公众号：腾讯混元","pageTitle":"텐센트 훈위안은 HyOCR-1.5를 출시했습니다: 6.37배 빠른 추론을 제공하는 종단 간 OCR 대형 모델 풀스택 오픈소스 - Aioga AI 뉴스","description":"텐센트 훈위안은 완전 오픈 소스 학습, 추론, 모델 가중치를 갖춘 엔드 투 엔드 OCR 대형 모델 분야에서 최초의 전문가 모델인 HyOCR-1.5를 출시했습니다. 1B 매개변수만으로 8개 이상의 텍스트 중심 작업을 다룹니다. DFlash 추측적 디코딩 프레임워크를 도입하여 트랜스포머에서 6.37× 가속, vLLM에서 2.","url":"https://www.aioga.com/ko/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:04.709Z"},"es":{"title":"Tencent Hunyuan lanzó HyOCR-1.5: un código abierto de gran tamaño OCR de extremo a extremo con una inferencia 6,37 veces más rápida","summary":"Tencent Hunyuan lanzó HyOCR-1.5, el primer modelo experto en el campo de grandes modelos OCR de extremo a extremo para entrenamiento, inferencia y ponderación de modelos completamente open source. Con solo 1B parámetros, cubre más de 8 tareas centradas en texto. Se introdujo el marco de decodificación especulativa DFLASH, logrando una aceleración del 6,37× bajo Transformers y del 2,14× bajo vLLM, con una inferencia de extremo a extremo que alcanzó 1,408 segundos por página. Soporta resolución 4K y ventanas contextuales de 128K, y amplía OCR de bajo recurso (331 idiomas), reconocimiento de escrituras antiguas y capacidades de preguntas y respuestas multiimagen mediante Flujo de Datos Agente. En OmniDocBench v1.6, se situó en primer lugar de ida a cara con una puntuación de 94,74.","category":"Modelos","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan lanzó HyOCR-1.5: un código abierto de gran tamaño OCR de extremo a extremo con una inferencia 6,37 veces más rápida - Aioga Noticias de IA","description":"Tencent Hunyuan lanzó HyOCR-1.5, el primer modelo experto en el campo de grandes modelos OCR de extremo a extremo para entrenamiento, inferencia y ponderación de modelos completame","url":"https://www.aioga.com/es/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:04.880Z"},"fr":{"title":"Tencent Hunyuan a publié HyOCR-1.5 : un gros modèle OCR de bout en bout open-stack open-source avec une inférence 6,37x plus rapide","summary":"Tencent Hunyuan a publié HyOCR-1.5, le premier modèle expert dans le domaine des grands modèles OCR de bout en bout pour l’entraînement, l’inférence et la pondération des modèles entièrement open source. Avec seulement 1B de paramètres, il couvre plus de 8 tâches centrées sur le texte. Introduction du cadre de décodage spéculatif DFlash, atteignant une accélération de 6,37 × sous Transformers et une accélération de 2,14 × sous vLLM, avec une inférence de bout en bout atteignant 1,408 seconde par page. Prend en charge la résolution 4K et les fenêtres contextuelles 128K, et étend l’OCR à faible ressources (331 langues), la reconnaissance des anciens scripts et les capacités de questions-réponses multi-images via Agentic Data Flow. Sur OmniDocBench v1.6, il s’est classé premier d’un bout à l’autre avec un score de 94,74.","category":"Modèles","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan a publié HyOCR-1.5 : un gros modèle OCR de bout en bout open-stack open-source avec une inférence 6,37x plus rapide - Aioga Actualités IA","description":"Tencent Hunyuan a publié HyOCR-1.5, le premier modèle expert dans le domaine des grands modèles OCR de bout en bout pour l’entraînement, l’inférence et la pondération des modèles e","url":"https://www.aioga.com/fr/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:04.834Z"},"de":{"title":"Tencent Hunyuan veröffentlichte HyOCR-1.5: ein End-to-End-OCR-Großmodell-Full-Stack-Open-Source-Modell mit 6,37-fach schnellerer Inferenz","summary":"Tencent Hunyuan veröffentlichte HyOCR-1.5, das erste Expertenmodell im End-to-End-Bereich der OCR-Großmodelle, das vollständig als Open Source, Training, Inferenz und Modellgewichtung dient. Mit nur 1B-Parametern deckt er mehr als 8 textzentrierte Aufgaben ab. Einführung des spekulativen Dekodierungsrahmens DFlash, das eine Beschleunigung von 6,37× unter Transformers und eine Beschleunigung von 2,14× unter vLLM erreicht, wobei die End-to-End-Inferenz 1,408 Sekunden pro Seite erreicht. Unterstützt 4K-Auflösung und 128K-Kontextfenster und erweitert ressourcenarme OCR (331 Sprachen), antike Skripterkennung und Multi-Image-Q&A-Funktionen durch Agentic Data Flow. Auf OmniDocBench v1.6 belegte es mit 94,74 den ersten Platz von End-to-End.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan veröffentlichte HyOCR-1.5: ein End-to-End-OCR-Großmodell-Full-Stack-Open-Source-Modell mit 6,37-fach schnellerer Inferenz - Aioga KI-News","description":"Tencent Hunyuan veröffentlichte HyOCR-1.5, das erste Expertenmodell im End-to-End-Bereich der OCR-Großmodelle, das vollständig als Open Source, Training, Inferenz und Modellgewicht","url":"https://www.aioga.com/de/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:11.141Z"},"pt-BR":{"title":"A Tencent Hunyuan lançou o HyOCR-1.5: um código aberto de grande formato full-stack de OCR de ponta a ponta com inferência 6,37x mais rápida","summary":"A Tencent Hunyuan lançou o HyOCR-1.5, o primeiro modelo especialista no campo de grandes modelos OCR de ponta a ponta para treinamento, inferência e ponderação de modelos totalmente open source. Com apenas 1B de parâmetros, ele cobre mais de 8 tarefas centradas em texto. Introduziu o framework de decodificação especulativa DFlash, alcançando 6,37× de aceleração sob Transformers e 2,14× sob vLLM, com inferência de ponta a ponta chegando a 1,408 segundos por página. Suporta resolução 4K e janelas de contexto 128K, além de expandir OCR de baixo recurso (331 idiomas), reconhecimento de escritas antigas e capacidades de perguntas e respostas multiimagem por meio do Fluxo de Dados Agente. No OmniDocBench v1.6, ficou em primeiro lugar de ponta a ponta com uma pontuação de 94,74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"A Tencent Hunyuan lançou o HyOCR-1.5: um código aberto de grande formato full-stack de OCR de ponta a ponta com inferência 6,37x mais rápida - Aioga Notícias de IA","description":"A Tencent Hunyuan lançou o HyOCR-1.5, o primeiro modelo especialista no campo de grandes modelos OCR de ponta a ponta para treinamento, inferência e ponderação de modelos totalment","url":"https://www.aioga.com/pt-BR/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:09.313Z"},"ru":{"title":"Tencent Hunyuan выпустил HyOCR-1.5: сквозно-конечную OCR-модель с полноценным открытым исходным кодом и в 6,37 раза быстрее вывода","summary":"Tencent Hunyuan выпустила HyOCR-1.5 — первую экспертную модель в области сквозной OCR с полным открытым исходным кодом обучения, вывода и взвешивания моделей. С параметрами 1B он охватывает более 8 текстовых задач. Введён фреймворк спекулятивного декодирования DFlash, достигающий ускорения 6,37× в Transformers и 2,14× ускорения при vLLM, при этом сквозная инференция достигает 1,408 секунды на страницу. Поддерживает разрешение 4K и 128K контекстных окна, а также расширяет малоресурсный OCR (331 язык), распознавание древних скриптов и возможности многообразных вопросов и ответов через Agentic Data Flow. В OmniDocBench v1.6 он занял первое место от первого до конца с результатом 94.74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan выпустил HyOCR-1.5: сквозно-конечную OCR-модель с полноценным открытым исходным кодом и в 6,37 раза быстрее вывода - Aioga Новости ИИ","description":"Tencent Hunyuan выпустила HyOCR-1.5 — первую экспертную модель в области сквозной OCR с полным открытым исходным кодом обучения, вывода и взвешивания моделей. С параметрами 1B он о","url":"https://www.aioga.com/ru/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:10.005Z"},"ar":{"title":"أصدرت Tencent Hunyuan HyOCR-1.5: نموذج OCR كبير من البداية إلى الطرف، مفتوح المصدر بنظام OCR متكامل مع استدلال أسرع بمقدار 6.37 مرة","summary":"أصدرت Tencent Hunyuan HyOCR-1.5، وهو أول نموذج خبير في مجال نماذج OCR الكبيرة من البداية إلى النهاية يدعم التدريب مفتوح المصدر بالكامل، والاستدلال، ووزن النماذج. مع 1B معلمة فقط، تغطي أكثر من 8 مهام نصية. تم تقديم إطار فك الترميز التكهني DFlash، محققا تسارعا بنسبة 6.37× تحت المحولات وتسارع 2.14× تحت تقنية vLLM، مع استنتاج من طرف إلى طرف يصل إلى 1.408 ثانية لكل صفحة. يدعم دقة 4K ونوافذ سياق 128K، ويوسع إمكانيات OCR منخفضة الموارد (331 لغة)، والتعرف على السكريبتات القديمة، وإمكانية الأسئلة والأجوبة المتعددة من خلال تدفق البيانات الوكيلي. في OmniDocBench v1.6، احتل المركز الأول من البداية إلى النهاية بدرجة 94.74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"أصدرت Tencent Hunyuan HyOCR-1.5: نموذج OCR كبير من البداية إلى الطرف، مفتوح المصدر بنظام OCR متكامل مع استدلال أسرع بمقدار 6.37 مرة - Aioga أخبار الذكاء الاصطناعي","description":"أصدرت Tencent Hunyuan HyOCR-1.5، وهو أول نموذج خبير في مجال نماذج OCR الكبيرة من البداية إلى النهاية يدعم التدريب مفتوح المصدر بالكامل، والاستدلال، ووزن النماذج. مع 1B معلمة فقط، ت","url":"https://www.aioga.com/ar/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:10.899Z"},"hi":{"title":"Tencent Hunyuan ने HyOCR-1.5 जारी किया: 6.37x तेज अनुमान के साथ एक एंड-टू-एंड OCR बड़ा मॉडल फुल-स्टैक ओपन-सोर्स","summary":"Tencent Hunyuan ने HyOCR-1.5 जारी किया, जो पूरी तरह से ओपन सोर्स प्रशिक्षण, अनुमान और मॉडल वेटिंग के लिए एंड-टू-एंड OCR बड़े मॉडल क्षेत्र में पहला विशेषज्ञ मॉडल है। केवल 1B मापदंडों के साथ, यह 8 से अधिक पाठ-केंद्रित कार्यों को कवर करता है। DFlash सट्टा डिकोडिंग फ्रेमवर्क पेश किया, ट्रांसफॉर्मर के तहत 6.37× त्वरण और vLLM के तहत 2.14× त्वरण प्राप्त किया, जिसमें एंड-टू-एंड अनुमान प्रति पृष्ठ 1.408 सेकंड तक पहुंच गया। 4K रिज़ॉल्यूशन और 128K संदर्भ विंडो का समर्थन करता है, और Agentic डेटा प्रवाह के माध्यम से कम-संसाधन OCR (331 भाषाएँ), प्राचीन स्क्रिप्ट पहचान और बहु-छवि Q&A क्षमताओं का विस्तार करता है। OmniDocBench v1.6 पर, यह 94.74 के स्कोर के साथ एंड-टू-एंड पहले स्थान पर रहा।","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan ने HyOCR-1.5 जारी किया: 6.37x तेज अनुमान के साथ एक एंड-टू-एंड OCR बड़ा मॉडल फुल-स्टैक ओपन-सोर्स - Aioga AI समाचार","description":"Tencent Hunyuan ने HyOCR-1.5 जारी किया, जो पूरी तरह से ओपन सोर्स प्रशिक्षण, अनुमान और मॉडल वेटिंग के लिए एंड-टू-एंड OCR बड़े मॉडल क्षेत्र में पहला विशेषज्ञ मॉडल है। केवल 1B मापदंडो","url":"https://www.aioga.com/hi/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:12.241Z"},"it":{"title":"Tencent Hunyuan ha rilasciato HyOCR-1.5: un open source open-source di grande formato OCR full-stack end-to-end con inferenza 6,37x più veloce","summary":"Tencent Hunyuan ha rilasciato HyOCR-1.5, il primo modello esperto nel campo dei modelli OCR end-to-end per addestramento, inferenza e ponderazione dei modelli completamente open source. Con solo 1B parametri, copre più di 8 compiti incentrati sul testo. Introdotto il framework di decodifica speculativa DFLASH, raggiungendo un'accelerazione del 6,37× sotto Transformers e del 2,14× con vLLM, con un'inferenza end-to-end che raggiunge 1,408 secondi per pagina. Supporta risoluzione 4K e finestre contestuali 128K, ed espande l'OCR a basse risorse (331 lingue), il riconoscimento degli antichi script e le capacità di Q&A multi-immagine tramite Agentic Data Flow. Su OmniDocBench v1.6, si è classificato primo end-to-end con un punteggio di 94,74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan ha rilasciato HyOCR-1.5: un open source open-source di grande formato OCR full-stack end-to-end con inferenza 6,37x più veloce - Aioga Notizie IA","description":"Tencent Hunyuan ha rilasciato HyOCR-1.5, il primo modello esperto nel campo dei modelli OCR end-to-end per addestramento, inferenza e ponderazione dei modelli completamente open so","url":"https://www.aioga.com/it/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:09.986Z"},"nl":{"title":"Tencent Hunyuan bracht HyOCR-1.5 uit: een end-to-end OCR large model full-stack open-source met 6,37x snellere inferentie","summary":"Tencent Hunyuan bracht HyOCR-1.5 uit, het eerste expertmodel in het end-to-end OCR-veld van grote modellen voor volledig open source training, inferentie en modelweging. Met slechts 1B-parameters dekt het meer dan 8 tekstgerichte taken. Het DFlash speculatieve decoderingsraamwerk werd geïntroduceerd, dat 6,37× versnelling bereikte onder Transformers en 2,14× versnelling onder vLLM, met end-to-end inferentie van 1,408 seconden per pagina. Ondersteunt 4K-resolutie en 128K-contextvensters, en breidt low-resource OCR (331 talen), ancient scriptherkenning en multi-image Q&A-mogelijkheden uit via Agentic Data Flow. Op OmniDocBench v1.6 stond het eerste van end-to-end met een score van 94,74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan bracht HyOCR-1.5 uit: een end-to-end OCR large model full-stack open-source met 6,37x snellere inferentie - Aioga AI-nieuws","description":"Tencent Hunyuan bracht HyOCR-1.5 uit, het eerste expertmodel in het end-to-end OCR-veld van grote modellen voor volledig open source training, inferentie en modelweging. Met slecht","url":"https://www.aioga.com/nl/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:09.519Z"},"tr":{"title":"Tencent Hunyuan, 6.37 kat daha hızlı çıkarıma sahip uçtan uca OCR büyük modelli tam yığın açık kaynaklı HyOCR-1.5'i piyasaya sürdü","summary":"Tencent Hunyuan, uçtan uca OCR büyük model alanında tamamen açık kaynak eğitim, çıkarım ve model ağırlıklandırma sunan ilk uzman model olan HyOCR-1.5'i piyasaya sürdü. Sadece 1B parametreyle, 8'den fazla metin odaklı görevi kapsar. DFlash spekülatif kod çözme çerçevesini tanıttı; Transformers ile 6.37× hızlanma, vLLM'de ise 2.14× hızlanma elde etti ve uçtan uca çıkarım sayfa başına 1.408 saniyeye ulaştı. 4K çözünürlük ve 128K bağlam pencerelerini destekler, düşük kaynaklı OCR (331 dil), eski script tanıma ve çok görüntülü Soru-Cevap yeteneklerini Ajanic Data Flow ile genişletir. OmniDocBench v1.6'da uçtan uca 94.74 puanla birinci oldu.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan, 6.37 kat daha hızlı çıkarıma sahip uçtan uca OCR büyük modelli tam yığın açık kaynaklı HyOCR-1.5'i piyasaya sürdü - Aioga AI Haberleri","description":"Tencent Hunyuan, uçtan uca OCR büyük model alanında tamamen açık kaynak eğitim, çıkarım ve model ağırlıklandırma sunan ilk uzman model olan HyOCR-1.5'i piyasaya sürdü. Sadece 1B pa","url":"https://www.aioga.com/tr/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:11.878Z"},"vi":{"title":"Tencent Hunyuan đã phát hành HyOCR-1.5: mã nguồn mở full-stack mô hình OCR lớn đầu cuối với suy luận nhanh hơn 6,37 lần","summary":"Tencent Hunyuan đã phát hành HyOCR-1.5, mô hình chuyên gia đầu tiên trong lĩnh vực mô hình lớn OCR đầu cuối để đào tạo, suy luận và trọng số mô hình mã nguồn mở hoàn toàn. Chỉ với 1 tỷ tham số, nó bao gồm hơn 8 tác vụ tập trung vào văn bản. Giới thiệu khung giải mã suy đoán DFlash, đạt được gia tốc 6,37× trong Transformers và tăng tốc 2,14× trong vLLM, với suy luận đầu cuối đạt 1,408 giây mỗi trang. Hỗ trợ độ phân giải 4K và cửa sổ ngữ cảnh 128K, đồng thời mở rộng khả năng OCR tài nguyên thấp (331 ngôn ngữ), nhận dạng tập lệnh cũ và khả năng hỏi đáp nhiều hình ảnh thông qua Luồng dữ liệu tác nhân. Trên OmniDocBench v1.6, nó xếp hạng đầu tiên từ đầu đến cuối với số điểm 94.74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan đã phát hành HyOCR-1.5: mã nguồn mở full-stack mô hình OCR lớn đầu cuối với suy luận nhanh hơn 6,37 lần - Tin tức AI Aioga","description":"Tencent Hunyuan đã phát hành HyOCR-1.5, mô hình chuyên gia đầu tiên trong lĩnh vực mô hình lớn OCR đầu cuối để đào tạo, suy luận và trọng số mô hình mã nguồn mở hoàn toàn. Chỉ với ","url":"https://www.aioga.com/vi/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:11.438Z"},"id":{"title":"Tencent Hunyuan merilis HyOCR-1.5: sumber terbuka full-stack model besar OCR end-to-end dengan inferensi 6,37x lebih cepat","summary":"Tencent Hunyuan merilis HyOCR-1.5, model ahli pertama di bidang model besar OCR end-to-end untuk pelatihan, inferensi, dan pembobotan model sumber terbuka sepenuhnya. Hanya dengan parameter 1B, ini mencakup lebih dari 8 tugas yang berpusat pada teks. Memperkenalkan kerangka kerja decoding spekulatif DFlash, mencapai akselerasi 6,37× di bawah Transformers dan akselerasi 2,14× di bawah vLLM, dengan inferensi end-to-end mencapai 1,408 detik per halaman. Mendukung resolusi 4K dan jendela konteks 128K, dan memperluas OCR sumber daya rendah (331 bahasa), pengenalan skrip kuno, dan kemampuan Tanya Jawab multi-gambar melalui Agentic Data Flow. Pada OmniDocBench v1.6, ia menempati peringkat pertama dari ujung ke ujung dengan skor 94.74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan merilis HyOCR-1.5: sumber terbuka full-stack model besar OCR end-to-end dengan inferensi 6,37x lebih cepat - Berita AI Aioga","description":"Tencent Hunyuan merilis HyOCR-1.5, model ahli pertama di bidang model besar OCR end-to-end untuk pelatihan, inferensi, dan pembobotan model sumber terbuka sepenuhnya. Hanya dengan ","url":"https://www.aioga.com/id/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:09.284Z"},"th":{"title":"Tencent Hunyuan เปิดตัว HyOCR-1.5: โอเพ่นซอร์สแบบฟูลสแตก OCR ขนาดใหญ่แบบ end-to-end พร้อมการอนุมานที่เร็วขึ้น 6.37 เท่า","summary":"Tencent Hunyuan เปิดตัว HyOCR-1.5 ซึ่งเป็นโมเดลผู้เชี่ยวชาญรุ่นแรกในฟิลด์โมเดลขนาดใหญ่ OCR แบบ end-to-end สําหรับการฝึกอบรมโอเพ่นซอร์ส การอนุมาน และการถ่วงน้ําหนักโมเดลอย่างเต็มรูปแบบ ด้วยพารามิเตอร์เพียง 1B จึงครอบคลุมงานที่เน้นข้อความเป็นศูนย์กลางมากกว่า 8 งาน เปิดตัวเฟรมเวิร์กการถอดรหัสการเก็งกําไร DFlash โดยบรรลุการเร่งความเร็ว 6.37× ภายใต้ Transformers และการเร่งความเร็ว 2.14× ภายใต้ vLLM โดยมีการอนุมานแบบ end-to-end ถึง 1.408 วินาทีต่อหน้า รองรับความละเอียด 4K และหน้าต่างบริบท 128K และขยาย OCR ทรัพยากรต่ํา (331 ภาษา) การจดจําสคริปต์แบบโบราณ และความสามารถในการถามตอบแบบหลายภาพผ่าน Agentic Data Flow ใน OmniDocBench v1.6 อยู่ในอันดับแรกแบบ end-to-end ด้วยคะแนน 94.74","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan เปิดตัว HyOCR-1.5: โอเพ่นซอร์สแบบฟูลสแตก OCR ขนาดใหญ่แบบ end-to-end พร้อมการอนุมานที่เร็วขึ้น 6.37 เท่า - ข่าว AI Aioga","description":"Tencent Hunyuan เปิดตัว HyOCR-1.5 ซึ่งเป็นโมเดลผู้เชี่ยวชาญรุ่นแรกในฟิลด์โมเดลขนาดใหญ่ OCR แบบ end-to-end สําหรับการฝึกอบรมโอเพ่นซอร์ส การอนุมาน และการถ่วงน้ําหนักโมเดลอย่างเต็มรูป","url":"https://www.aioga.com/th/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:12.287Z"},"pl":{"title":"Tencent Hunyuan wypuścił HyOCR-1.5: end-to-end model OCR z pełnym stosem open-source z 6,37 razy szybszym wnioskowaniem","summary":"Tencent Hunyuan wypuścił HyOCR-1.5, pierwszy ekspercki model w kompleksowym obszarze OCR dla dużych modeli w pełni otwartego oprogramowania, szkolenia, wnioskowania i ważenia modeli. Przy zaledwie 1B parametrach obejmuje ponad 8 zadań skoncentrowanych na tekście. Wprowadzono framework spekulatywnego dekodowania DFlash, osiągający przyspieszenie 6,37× pod Transformers i 2,14× przyspieszenie w vLLM, z end-to-end wnioskowaniem osiągającym 1,408 sekundy na stronę. Obsługuje okna kontekstowe w rozdzielczości 4K i 128K, a także rozszerza OCR o niskich zasobach (331 języków), rozpoznawanie starych skryptów oraz wieloobrazowe Q&A dzięki Agentic Data Flow. Na OmniDocBench v1.6 zajęła pierwsze miejsce end-to-end z wynikiem 94,74.","category":"模型更新","source":"公众号：腾讯混元","pageTitle":"Tencent Hunyuan wypuścił HyOCR-1.5: end-to-end model OCR z pełnym stosem open-source z 6,37 razy szybszym wnioskowaniem - Aioga Wiadomości AI","description":"Tencent Hunyuan wypuścił HyOCR-1.5, pierwszy ekspercki model w kompleksowym obszarze OCR dla dużych modeli w pełni otwartego oprogramowania, szkolenia, wnioskowania i ważenia model","url":"https://www.aioga.com/pl/news/cmrj4s89p05nhbilkljqel2d5/","contentTranslated":true,"sourceHash":"c2b53ee93724ea45","translatedAt":"2026-07-19T12:15:11.730Z"}}}}