{"id":46447,"date":"2025-11-25T19:03:14","date_gmt":"2025-11-25T11:03:14","guid":{"rendered":"https:\/\/www.1ai.net\/?p=46447"},"modified":"2025-11-25T19:03:14","modified_gmt":"2025-11-25T11:03:14","slug":"%e8%85%be%e8%ae%af%e6%b7%b7%e5%85%83-ocr-%e6%a8%a1%e5%9e%8b%e5%ae%a3%e5%b8%83%e5%bc%80%e6%ba%90%ef%bc%9a%e5%8f%82%e6%95%b0%e4%bb%85-1b%ef%bc%8c%e5%a4%9a%e9%a1%b9%e6%a0%b8%e5%bf%83%e8%83%bd%e5%8a%9b-so","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/46447.html","title":{"rendered":"OCR MODEL DECLARED OPEN SOURCE: PARAMETER 1B, MULTIPLE CORE CAPABILITIES SOTA"},"content":{"rendered":"

The news of November 25thTencent Hunyuan<\/a>Today announced the launch of the newOpen Source Model<\/a>\u00a0HunyuanOCR<\/strong>THE PARAMETER IS ONLY 1B AND IS BASED ON A MULTI-MODULAR STRUCTURE BASED ON A MULTI-INDUSTRY OCR APPLICATION LIST SOTA (NOTE: STATE OF THE ART)\u3002<\/p>\n

\"3ad98f4fj00t6a3zm000td000uf9m\"<\/p>\n

According to official sources, thanks to the conceptual design of the MMA \u201cend-to-end\u201d concept, the functions of the HunyuanOCR are best achieved by a single forward reasoning\u3002<\/p>\n

THE OCR EXPERT MODEL IS BASED ON A MULTIMODULAR STRUCTURE CONSISTING OF THREE MAIN COMPONENTS:Native resolution video encoder, self-adapted visual adapter and light quantified hybrid language model<\/strong>.<\/p>\n

Unlike other open-source OCR expert models or systems, the training and reasoning of the HunyuanOCR model is based on a whole-to-end paradigm, with robust end-to-end reasoning demonstrated through scaled-up application-oriented data combined with enhanced online learning\u3002<\/p>\n

The hybrid OCR has several core competencies that achieve SOTA effects, of which the OmniDocBench assessment of complex document resolution achieves the highest 94.1 pointsMore than Google's Gemini3-pro and so on<\/strong>; the word detection and recognition capability, in the benchmark of the self-built 9 major applications (documentation, art, street scene, handwritten, advertising, paper, screen-stopping, games, video), is a significant lead in the same open-source model and the commercial OCR model; on the OCRBench listTHE TOTAL SCORE WAS 860 POINTS, AND THE MODEL CONFIGURATION OF ONLY 1B TOTAL PARAMETER OBTAINED THE TOTAL PARAMETER, INCLUDING THE GENERAL VISUAL UNDERSTANDING MODEL 3B UNDER SOTA<\/strong>.<\/p>\n

IN SMALL-LANGUAGE TRANSLATION SKILLS, MIXED OCR SUPPORTS 14 HIGH-FREQUENCY SMALL-LANGUAGE TRANSLATIONS INTO CHINESE OR ENGLISH AND HAS WON THE ICDAR2025-END-TO-END DOCUMENT TRANSLATION SMALL-MODEL CHAMPION\u3002<\/p>\n

\"0c6983dfj00t6a3zm002td000u000m5m\"<\/p>\n

In terms of applications, HunyuanOCR supports the resolution of complex documents in a multilingual language, with a combination of text detection and recognition capabilities, and applications in such settings as paper field extraction, video subtitle recognition, photo translation, etc\u3002<\/p>\n

In terms of text detection and recognition, models perform well on scenes such as documents, art words, street scenes, handwritten writings, advertising, bills, screens, games, videos, etc\u3002<\/p>\n

Complex document resolution refers to the electronicization of a multilingual document scanned or image taken, specifically, the organization of text elements that appear in a picture in the reading order, the use of the Latex formulae, and the presentation of complex tables in HTML format\u3002<\/p>\n

In addition to the usual applications, there is a need for field extraction, video subtitle extraction and photo translation\u3002<\/p>\n

1 interest fields for common cards and instruments (e.g. name \/ address \/ unit, etc.) are analysed in standard json format\u3002<\/p>\n

2. Automation of subtitles of videos, including bilingual subtitles\u3002<\/p>\n

3. Photo-translator function, which supports 14 small languages for high frequency applications, including: German, Spanish, Turkish, Italian, Russian, French, Portuguese, Arabic, Thai, Vietnamese, Indonesian, Malay, Japanese, Korean, and Chinese\/English\u3002<\/p>\n

1AI with the following open source addresses:<\/p>\n