Hunyuan Big Model officially debuts and is open to the public through Tencent Cloud

According to Fast Technology today (September 7), at the 2023 Tencent Global Digital Ecosystem Conference, Tencent's Hunyuan Model was officially unveiled and announced to be open to the public through Tencent Cloud.

It is understood that Tencent Hunyuan Large Model is a general large language model developed by Tencent throughout the entire chain. It has a parameter scale of over 100 billion and a pre-trained corpus of over 2 trillion tokens. It has strong Chinese creation capabilities, logical reasoning capabilities in complex contexts, and reliable task execution capabilities.

From parameter-first to practical-first

It is worth noting that the Tencent Hunyuan Big Model is a practical big model that "comes from practice and goes to practice". More than 50 Tencent businesses and products, including Tencent Cloud, Tencent Advertising, Tencent Games, Tencent Financial Technology, Tencent Conference, Tencent Documents, WeChat Search, QQ Browser, etc., have been connected to the Tencent Hunyuan Big Model test and have achieved initial results.

It is understood that the Hunyuan big model will serve as the foundation of Tencent Cloud's MaaS service. Customers can not only call Hunyuan directly through the API, but also use Hunyuan as a base model to build exclusive applications for different industrial scenarios.

Tom Tong, senior executive vice president of Tencent Group and CEO of the Cloud and Smart Industries Group, said: "With big model generation technology at its core, artificial intelligence is becoming the key driving force for the next round of digital development, and it also brings new ideas for solving industry pain points. Big models need to be based on industry scenarios and integrated with enterprise data to release the greatest value."

Self-developed full-link technology

According to Jiang Jie, vice president of Tencent Group, Tencent's Hunyuan large model has been trained from scratch since the first token, and has mastered the full-link self-developed technology from model algorithms to machine learning frameworks to AI infrastructure.

Since 2021, Tencent has successively launched NLP sparse large models with hundreds of billions and trillions of parameters, breaking the three major CLUE list records and achieving new breakthroughs in Chinese comprehension capabilities.

At present, the application of large models in the industry is still limited, mainly concentrated in leisure scenarios with high fault tolerance and simple tasks. Tencent has carried out a series of self-developed innovations at the algorithm level to improve the reliability and maturity of the model.

To address the problem of large models being prone to "nonsense", Tencent optimized the pre-training algorithms and strategies, reducing the hallucinations of the Hunyuan large model by 30% to 50% compared to mainstream open source large models; through reinforcement learning methods, the model learned to identify trap problems; through positional encoding optimization, the processing effect and performance of very long texts were improved; and a new thinking chain strategy was proposed, allowing the large model to reason and make decisions based on actual application scenarios like humans.

In addition, Tencent has also developed its own machine learning framework Angel, which increases the training speed by 1 times and the inference speed by 1.3 times compared to the mainstream framework in the industry.

Thanks to the full-link self-developed technology, Tencent Hunyuan Big Model can understand the meaning of the context and has the ability to memorize long texts, and can smoothly conduct multiple rounds of conversations in professional fields. In addition, it can also create content such as literary creation, text summarization, role-playing, etc., fully understand the user's intentions, and give efficient and accurate responses with timeliness.

In the standard compliance test of CAICT's "Evaluation Methods for Large-Scale Pre-trained Model Technology and Applications", the Hunyuan Large Model was evaluated in 66 capability items, and received the highest score in the comprehensive evaluation of the two important areas of "model development" and "model capability". The Hunyuan Large Model performed well on the mainstream evaluation sets MMLU, CEval and AGI-eval, especially in the sub-items of science, college entrance examination questions and mathematics in Chinese.

Liu Yuanchun, president of Shanghai University of Finance and Economics, believes that: "With the help of full-link self-research, China will continue to accumulate talents and technologies related to large models, gradually form a systematic industrial chain, talent chain, technology chain and innovation chain, and finally find a Chinese path for the development of general artificial intelligence, helping us to make breakthrough progress in the innovation of digital technology."

Tencent fully embraces the big model

Jiang Jie said: "Our goal in developing big models is not to get high scores in evaluation, but to apply the technology to actual scenarios. Tencent will fully embrace big models."

At the conference, Jiang Jie demonstrated the actual application of multiple businesses such as Tencent Meeting, Tencent Docs, and Tencent Advertising after they were connected to the Tencent Hunyuan Big Model.

For example, Tencent Meeting has created an AI assistant based on the Hunyuan Big Model, which can complete complex tasks such as conference information extraction and content analysis with simple natural language commands, and can also generate intelligent summary minutes after the meeting. According to actual tests, the Hunyuan Big Model has achieved a high user adoption rate in many aspects such as command understanding, in-meeting Q&A, meeting summaries, and meeting to-do items.

In terms of document processing, Tencent Hunyuan's large model supports dozens of text creation scenarios and has been applied in the smart assistant function launched by Tencent Documents. At the same time, Hunyuan can also generate standard format text with one click, master hundreds of Excel formulas, support natural language generation functions, and generate charts based on table content. These functions are currently in the internal testing stage and will be open to users after they are mature.

In advertising business scenarios, Tencent Hunyuan Big Model supports intelligent advertising material creation, can adapt to industry and regional characteristics, meet the needs of thousands of people, and achieve the natural integration of text, pictures, and videos. In addition, based on the capabilities of Hunyuan Big Model, advertising smart shopping guides can help merchants improve service quality and efficiency in scenarios such as corporate WeChat.

It is understood that in June this year, Tencent Cloud launched the Model as a Service (MaaS) solution, providing one-stop industry big model services covering model pre-training, model fine-tuning, intelligent application development, etc.

Recently, Tencent Cloud has also fully integrated more than 20 mainstream models such as Llama 2 and Bloom, and like Hunyuan, they all support direct deployment and calling. Customers can build their own exclusive industry models based on Hunyuan or open source models according to actual needs.