Tencent Hunyuan upgrades model matrix, launching 256k long text model on the cloud​

The implementation of large models is accelerating, and "industrial practicality" has become the development consensus.

On May 17, 2024, Tencent Cloud Generative AI Industry Application Summit was held in Beijing, announcing large model research and development, application products series progress.

Tencent’s Hunyuan large model capabilities continue to upgrade. Multiple versions of models hunyuan-pro, hunyuan-standard, and hunyuan-lite are open to the outside world through Tencent Cloud to meet the needs of enterprise customers and developers in different situations. The model requirements in the scenario are , and the optimal cost-effective model solution is implemented .

Tencent Cloud released three major tools: large model knowledge engine, image creation engine, and video creation engine, creating a native tool chain in the era of large models, simplifying data access and model fine-tuning through PaaS services , application development process, helping enterprises to use large models to develop AI native applications more efficiently and simply, and quickly access production scenarios.

Taking "industrial practicality" as the core strategy for the development of large models, Tang Daosheng, senior executive vice president of Tencent Group and CEO of cloud and smart industry group, said that through Create high-performance models, efficient tool platforms, highly agile scenario applications, highly available computing infrastructure, and strong security model environments to build AI that is closest to the industry.

（

Tencent Group Senior Executive Vice President, Cloud and Intelligence Industry Group CEO Tang Daosheng)

##Tencent Hunyuan Open Cloud 256k long text model, opening up the intelligent agent ecosystem

Powerful general-purpose large models and low-threshold development tools can improve developer productivity and promote the development of large model ecosystems.

At this industry summit, Tencent Hunyuan introduced its multi-size LLM model matrix. Among them, the largest model has been expanded to trillions of parameter scale, and is in 1B, 3B , 7B, 13B and other different parameter amounts, there are layouts.

The upgraded Tencent Hunyuan is the first in China to adopt the mixed expert model (MoE) structure. The overall performance of the model has been improved by 50% compared with the previous generation, and some Chinese capabilities

Already tied with GPT-4, with great improvement in answering performance of "current" questions, mathematics, reasoning and other abilities. On Tencent Cloud, Hunyuan large model also provides models of multiple sizes such as hunyuan-pro with trillion parameters, hunyuan-standard with hundreds of billions of parameters, and hunyuan-lite with billions of parameters. The service is now fully open to enterprises and individual developers.

(Tencent’s hybrid model service is open to the outside world through Tencent Cloud)

Among them, hunyuan-standard has recently launched a long text model that supports 256k ultra-long context windows. It has the ability to process ultra-long texts of more than 380,000 characters at a time. It is good at reading comprehension and reading of long documents. It has demonstrated powerful performance in large-scale data analysis and can provide strong work support for professionals in finance, medical, education, travel and other industries, significantly improving work efficiency.

In terms of multi-modal capabilities, the Hunyuan large model also continues to be iteratively upgraded. In the field of raw graphics, Tencent Hunyuan Wenshengtu infrastructure has been fully upgraded to the same DiT architecture as Sora, supporting bilingual input and understanding in Chinese and English, with multi-round drawing capabilities, and leading domestic evaluation results; in the field of raw video, Tencent Hunyuan Supports various video generation capabilities such as text video, picture video, picture and text video, video video, etc.

has supported 16s video generation; at the 3D level, Tencent Hunyuan has Lay out text/pictures to produce 3D, and it only takes 30 seconds to generate a 3D model from a single picture. According to Sullivan’s evaluation results, Tencent Hunyuan’s general basic capabilities and professional application capabilities are both in the leading echelon of domestic large models and higher than the average of international large models. A report from the authoritative evaluation agency SuperCLUE also shows that Tencent's Hunyuan large model ranks in the first echelon of domestic large models, taking a leading position in both basic and scene applications, and is located in the Excellent Leaders Quadrant.

At the meeting, Jiang Jie, Vice President of Tencent Group, announced that Tencent’s Hunyuan model will embrace open source. Previously, the large model of Hunyuan Wenshengtu has been fully open sourced and received the attention of over a thousand developers on Github in just 3 days. Tencent's hybrid MoE models of various sizes will also be open sourced to the outside world, and can support diverse deployment scenarios such as mobile phones, PCs, clouds/data centers, etc.

（Tencent Group Vice President Jiang Jie）

As a practical large model, the Hunyuan large model has been tested in more than 600 Tencent internal businesses and scenarios, and continues to iterate in Tencent's rich ecosystem. Based on the Hunyuan large model, WeChat Reading has launched new functions such as AI book questioning and AI outline, which greatly improves users' reading efficiency and experience. The Tencent customer service team upgraded the intelligent customer service system based on the Hunyuan large model, and created an original fine-tuned model in the vertical field of intelligent customer service, which greatly improved the accuracy of understanding the intention of intelligent dialogue and the smoothness of multiple rounds of questions and answers. Compared with the traditional small model below The accuracy has increased by 38%, and its manual customer service assistant has been applied in multiple game customer service scenarios, with average daily user requests reaching 1.5 million times. Tencent Meeting is based on the AI ​​assistant launched by Hunyuan, which can instantly answer questions inside and outside the meeting, greatly improving meeting efficiency. In the past four months, the number of daily calls to Tencent’s conference AI assistant has increased 20 times. Collaborative SaaS products such as Enterprise WeChat and Tencent Documents are also fully integrated into Tencent Hunyuan. Tencent Advertising launched a one-stop AI advertising creative platform based on Tencent Hunyuan - Tencent Advertising Miaoshi to help improve the efficiency of advertising production and delivery.

Jiang Jie said that externally, Tencent Hunyuan will also open up the intelligent agent ecosystem and launch the one-stop AI intelligent agent creation and distribution platform "Tencent Yuanqi". In the future, users can not only create exclusive AI agents on the platform and use Tencent's official plug-ins and knowledge bases, but also distribute these agents to QQ, WeChat customer service, Tencent Cloud and other channels with one click.

（「Tencent Yuanqi」Official website is open for trial applicationhttps://www.php.cn/link/9afa24d3da745fd5606e7d710a0763eb）

Tencent Cloud releases three major AI large model engines to create a knowledge service application in 5 minutes

With large model technology as the core , Artificial intelligence has become a key driving force for the digital development of enterprises. Research shows that more than 60% of Chinese companies plan to deploy generative AI in the next 12 to 24 months.

But how to identify the correct scenario, deploy quickly, and shorten the distance from basic model to industrial application?

In the past year, Tencent has discovered that the industry’s demand for models is constantly changing in the process of serving industrial customers. On the one hand, with the diversification of industrial information carriers, the model needs to process not only simple text, but also pictures, videos and other information. The capability competition of large models has expanded from a single text-based model to multi-modal capabilities such as text-based pictures, text-based videos, picture-based pictures, and picture-based videos. On the other hand, in the environment of cost reduction and efficiency improvement, enterprises have higher requirements for "cost-effectiveness" and expect to use simpler large-model tools to accelerate application development, achieve rapid production, and meet a sustainable input-output ratio.

In order to better solve these needs, Tencent Cloud brand newlaunchesThe large model native tool chain uses three PaaS products-"Large Model Knowledge Engine", "Large Model Image Creation Engine" and "Large Model Video Creation Engine" to help enterprises in the field of knowledge creation Improve quality and efficiency in services, image and video creation.

(Tencent Cloud releases three major AI engine tools to lower the threshold for model application)

Among them, the large-model knowledge engine focuses on enterprise knowledge service scenarios and is based on the RAG (Retrieval Enhanced Generation) technical architecture. It integrates OCR document parsing and vector Retrieval, large language model, multi-modal large model and other technologies create a "low threshold" and "high performance" model application development platform for enterprises. Through "modular" application templates, enterprises can use natural Language,You can develop a knowledge service application in 5 minutes,quickly implement it in various business scenarios that connect people, such as customer service marketing and corporate knowledge communities.

At present, Tencent Cloud's large model knowledge engine has been implemented in many industries such as government affairs, finance, education, travel, and retail. In the financial industry, Yuanxin Huibao has developed an efficient people-benefiting think tank for insurance agents. Assisted by large-scale model technology, it automatically generates product knowledge questions and answers and soothing words, achieving a 50% per capita efficiency increase. In the education industry, Henan Digital Education Development Co., Ltd. uses a knowledge engine to import millions of primary and secondary school textbook documents in Henan Province, conducts knowledge sorting and configuration, and creates a 7×24-hour large-model knowledge teaching assistant.

Within Tencent, many SaaS applications are upgraded based on the knowledge engine. In the customer service scenario, Qidian customer service large-model text robot is connected to the large-model multi-round task engine to perform tasks such as bill inquiry, returns and exchanges, and the configuration cost is 50% lower than that of traditional text robots. In the digital human service scenario, after the digital human is connected to the large model knowledge engine, it can better understand and identify user intentions, and use the large model to generate more professional and personalized answers. In the enterprise knowledge service scenario, Tencent Enjoy combines the knowledge engine to provide "intelligent writing and generation" capabilities on the knowledge production side and "intelligent question and answer" capabilities on the knowledge consumption side, allowing enterprise employees to produce and learn knowledge Be more efficient and improve organizational capabilities.

In addition to the large model knowledge engine, the image and video creation engine will comprehensively improve the efficiency of material generation through large models. The "Image Creation Engine" is based on Tencent Hunyuan's self-developed image creation underlying model and outputs high-qualityAI image generation and editing capabilities to provide corporate customers with AI photography, line drawing, image stylization and other capabilities. For example, in the design scenario, corporate customers use the "line draft drawing" function to upload product line draft design drawings and then, and quickly generate physical design drawings through prompt words and parameter settings. , significantly shortening the creation and production cycle.

"Video creation engine" is based on multi-modal algorithm technology, outputs high-quality video generation and processing capabilities, and provides video translation, video stylization, and canvas expansion and many other functions. Faced with the needs of enterprises to go overseas, "Video Translation" helps corporate customers translate original videos into multiple languages ​​for video output with one click, quickly put them into overseas markets, and seize sales opportunities.

（Vice President of Tencent Cloud, Head of Tencent Cloud Intelligence, Head of Youtu Lab, and Head of Tencent Qidian Wu Yunsheng)

## Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence, said that Tencent Cloud has created large-scale models based on the actual needs of the industry. Era's native tool chain relies on three major AI large model engine tools to simplify the processes of data engineering, model fine-tuning, and application development, helping enterprises to use large models more efficiently and conveniently.

## Computing power and security dual base upgrade to escort the development of generative AI

Generative AI drives "intelligence emergence", bringing growth opportunities to enterprises, and also brings

newsecurity challenges. In the process of industrial practice, Tencent found that two major obstacles for enterprises to embrace generative AI are the shortage of computing resourcesandsecurityworries.

Security compliance is the bottom line for enterprises to apply

artificial intelligence technology. Based on more than 20 years of security technology accumulation, Tencent Security has upgrade launched a systematic security solution for AIGC scenarios.

In terms of data security, Tencent Security has launched a full-link data security solution through key management systems, bastion machines, data security governance centers (API security monitoring) and other tools , escorting data security throughout the entire life cycle of enterprise model training, fine-tuning, release, and operation, helping enterprises protect sensitive data and ensure data collection security and compliance.

In terms of content security, content generated by large models often encounters common risks such as false information, content infringement, induction risks, and personal privacy. Tencent Cloud AIGC content compliance solution is solved through five major service systems: expert services, data services, copyright services, machine review services, and CEM (customer experience management) services.

#model training, content generation, and post-operation operations of AIGC applications and other

content security challenges.

At present, Tencent Yuntianyu has escorted multiple AIGC business formats, covering scenarios such as AI Q&A, digital people, creative assistants, Wensheng diagrams, code generation, entertainment and social networking, and AI customer service. ######

In terms of computing power, Tencent Cloud provides one-stop AI infrastructure for industry training of large models. Through its self-developed Xingmai network 3.2T communication bandwidth and unified access layer capabilities, Tencent Cloud has created a computing cluster that can support parallel computing of more than 100,000 cards and is compatible with a variety of GPU ecosystems. Tencent Cloud has also launched China's first AI-native vector database, which supports up to 100 billion vector scales. It is the first domestic product to pass the vector database capability evaluation of the Academy of Information and Communications Technology.

##Generative AI ecological plan is released to build a prosperous ecosystem and drive industrial intelligent change

The industrial implementation of large models is a vast market, and it is also a complex process that requires the entire industry chain of large model manufacturers, physical industries, and ecological partners. Since 2023, Tencent Cloud has worked closely with 1,500 partners, relying on leading and rich generative AI products, serving more than 20,000 enterprise customers, and initially built an ecosystem around generative AI products.

At this summit, Tencent Cloud officially launched the generative AI ecological plan. Yang Chen, vice president of Tencent Cloud and head of industrial ecological cooperation, said that Tencent Cloud focuses on strengthening generative AI technology and platform base, and will open up platform capabilities and services and provide technical and market support in the future. Join thousands of solution providers, cultivate thousands of service providers and tens of thousands of agents, jointly promote the penetration of generative AI technology into the entire industry chain, and accelerate the intelligent upgrading of the industry.

(Tencent Cloud joins 17 partners to release generative AI ecological plan)

##In terms of capability opening, Tencent Cloud will open a full range of generative AI products to solution providers, as well as PaaS, aPaaS, iPaaS and other platforms and atomic capabilities to help partners on the application side. Develop differentiated and competitive solutions. In terms of service opening, Tencent Cloud will focus on cultivating thousands of professional service partners for generative AI products. Through systematic competency certification, partners will form a full-process service capability matrix to improve quality delivery for industry implementation. In terms of technical and market support, Tencent Cloud provides partners with special testing funds for seven major generative AI core products such as knowledge engine, vector database, and digital intelligence, and arranges for 100 large model architects, product experts, algorithm experts, and data experts. Experts accompany you to help partners accelerate implementation and replication in customer scenarios. At the same time, a generative AI marketing multiplier fund has been set up to effectively help partners develop markets.

At the meeting, Tencent Cloud also jointly released the "Generative AI Industry Implementation Path Research Report" (hereinafter referred to as the "Report") in conjunction with Gartner, providing enterprises with a matrix of generative AI application scenarios. and generative AI application implementation roadmap to help enterprises solve challenges such as scenario value and implementation feasibility.

Tang Daosheng said that facing the intelligent future, Tencent will always adhere to the strategic direction of "industrial practicality" and insist on using technology to solve practical problems. It will also adhere to ecological openness and cooperate with the industry. and partners to help the industry embrace the intelligent future

The above is the detailed content of Tencent Hunyuan upgrades model matrix, launching 256k long text model on the cloud​. For more information, please follow other related articles on the PHP Chinese website!