On June 19, Tencent Cloud announced for the first time the progress of research and development of industry models.。Tencent said on its WeChat public account that it will help enterprises build exclusive large models and intelligent applications, and will create industry large model selection stores, comprehensively lower the landing threshold, and provide MaaS (Model as a Service) one-stop service for enterprises.。

What are the MaaS one-stop services of Tengxun Cloud？

Tencent's MaaS one-stop service for enterprises includes large model selection, customization and landing.。Specifically, Tencent will rely on the Tencent Cloud TI platform, combined with the past on the travel, finance, government, education and other industries demand insights, packaged into a large model of the industry solutions to provide customers.。Enterprises can also add their own unique scenario data to these built-in industry models to quickly generate their own custom large models.。

Tencent will also open a large model toolkit that includes data labeling, training, evaluation, testing and deployment services for enterprises to customize large models with different parameters and specifications based on the original model.。Among them, the "Tai Chi Angel" upgraded by TI-ACC will provide excellent training and reasoning acceleration for users to train large models.。

In terms of large model computing power, Tencent is also well prepared this time。

One is the training framework AngelPTM developed by Tengxun Cloud.。The training framework completed the first trillion-parameter large model training in October last year and reduced the training time by 80% under the same data set.。

The second is a new generation of high-performance computing clusters HCC。In April this year, Tencent Cloud released a new generation of high-performance computing cluster HCC-HCCPNV5 for large model training.。HCCPNV5 uses the Star Sea self-developed server, equipped with NVIDIA H800 Tensor Core GPU, providing 3.2Tbps RDMA network, the overall performance of the cluster is up to 3 times higher than that of the previous generation HCCPNV4h, which further improves the training efficiency of large model clusters under the premise of GPU parallel computing.。HCCPNV5 is good at compute-intensive and massive data processing, and its application scenarios include distributed AI training (mainly including AIGC large models, autonomous driving, NLP, recommendation systems, etc.) and scientific computing (mainly including pharmaceuticals, protein policies, molecular dynamics, etc.). It is one of the most powerful large model computing clusters in China.。

The third is a vector database for AI operations.。The database can efficiently process unstructured data such as images, audio and text。Daily retrieval can reach hundreds of billions of levels, single index processing scale of billions of levels, throughput of 100,000 levels, can provide efficient retrieval support for large models.。

In terms of cost, Tencent Cloud TI platform enables low-cost landing for industry scenarios。It is reported that Tencent cloud TI platform on the basis of a small amount of computing power, improve the effect of specific tasks, such as intelligent customer service scenarios, training performance can be improved by 10 times, training costs decreased by 90%。

This time, Tencent's many head SaaS products, such as Tencent Enterprise Point, Tencent Conference, AI Code Assistant, etc. have also been upgraded.。

● Tencent Conference: The upgraded AI assistant can coordinate participants' time, meeting minutes, intelligent information extraction, intelligent sound adjustment and other functions.。

● Tencent Enterprise Point: It is reported that the enterprise point analysis AI assistant has five core competencies, namely: data understanding, indicator dismantling, language query, AI reasoning, conclusion extraction.。Based on these five capabilities, the AI assistant can quickly generate visual charts on business data and generate analysis reports as needed.。Secondly, the intelligent customer service inside can call the business system to provide real-time data, thus providing more accurate and detailed answers.。

● Tencent Cloud AI Code Assistant: supports code completion and code correction based on large model capabilities, and assists developers in improving development quality.。AI code assistant supports C / C + +, JavaScript, Go, Python and other languages and common open frameworks, as well as VS Code and other mainstream IDE。

In addition, Tencent Cloud will also provide enterprises with full escort optimization and localization training services, security experience and Tianyu wind control capabilities, and provide AIGC full-link content compliance solutions to ensure that large models are credible, reliable and available.。

Why the Big Model of Industry Solutions？

It is worth noting that, unlike other major technology companies that first announced the general language model, Tencent is the first to target the industry model.。

In fact, the reason behind this is that enterprises have this demand.。Under the wave of AI, many enterprises do not want to miss the opportunity of AI, and AI can also help enterprises improve efficiency in a variety of work scenarios。However, in contrast to general-purpose large models such as ChatGPT, industry-specific large models are basically focused on specialized segments and can be described as industrial versions of ChatGPT。

But for traditional or digital startups, the training data for large models is often unfiltered and processed, mostly containing publicly available information such as errors, rumors and biases.。If you "feed" the big model directly, not only does the big model not work well, but it also wastes a lot of cost.。In addition, large model training requires high computing, storage, and talent resources, among other things.。The larger the model, the higher the cost of training, so companies also tend to choose the right large model at a reasonable cost.。

Zhou Ming, founder and CEO of Chinese model provider Lanzhou Technology, has said that if Chinese companies want to develop models and solutions similar to ChatGPT, they need to consider the resources and labor costs of graphics processors.。

Tencent can solve the above problems by providing MaaS one-stop service for enterprises.。From building models to "feeding data" to "after-sales service," all of them are covered.。Moreover, the same is "feed data," Tencent can provide enterprises with more accurate data "food," and at a lower cost as soon as possible to build a more professional and accurate exclusive large model.。Product good + low cost, Tencent a move, that is, accurately stepped on the enterprise's pain point。

Tencent also summarizes the five advantages of its industry big model。

According to Tencent, it has provided more than 50 large model industry solutions for more than 10 industries, including finance, culture and tourism, government affairs, media and education, and its customers include CCTV and Shanghai University.。And Tencent Cloud has launched cooperation with 22 enterprises to build large models of the industry, and 17 ecological partners jointly launched the "Tencent Cloud Industry Large Model Ecological Plan."。

Tang Daosheng, senior executive vice president of Tencent Group and CEO of Cloud and Smart Industry Group, said, "Today, we are once again standing at the singularity of the digital technology revolution, the big model is just the beginning, and the integration of AI and industry will bloom a creative future.。"

CICC: Tencent prioritizes ToB's approach to AI, reflecting its pragmatic spirit

In response to the release of Tencent's large model, CICC released a research report saying that for the hot AI field, Tencent gave a distinctive answer: relying on Tencent Cloud TI platform to build industry large model selection stores - internal "cut" out for different industries of the vertical model, enterprises can quickly fine-tune the generation of exclusive models and customized services, the case covers finance, media, culture and tourism, government affairs, education and other industries。This just confirms Ma Huateng's previous statement: "For the industrial revolution, taking out the light bulb a month earlier is not so important in a long time span.。The key is still to get the underlying algorithms, computing power and data down to earth, and more importantly, the scene landing。"The bank believes that Tencent's priority to use ToB's thinking to enter AI is a manifestation of its pragmatic spirit, and it is expected to become one of the first companies in the industry to touch the commercialization of AI - Tencent Cloud itself has the accumulation of B-end customers, while for B-end users, the industry model is more exclusive and efficient, which is more conducive to solving the current pain points of enterprises in the short term, and enterprises are more willing to pay.。

In addition, the bank believes that the integration of AI and applications is expected to have more explicit and excellent cases, thus further consolidating the advantages of Tencent Cloud SaaS products, revitalizing Tencent Cloud's SaaS ecosystem, and leveraging its potential for commercialization.。The overall progress of Tencent's AI model: internal and external repair。As an important infrastructure for AI, Tencent has accumulated a large model of mixed yuan (released last year), a large language model of WeChat, and a more segmented large model of Tencent's cloud industry.。According to the report, Tencent invested in three AI big model start-ups (MiniMax, Shenyan Technology, Lightyear) in June this year, with a view to leveraging external technical forces to build capabilities in the AI big model field.。The bank believes that the overall progress of Tencent's AI model continues Tencent's style - both internal and external, with internal business research and development, collaboration and external investment layout to work together on AI long-term business.。