After OpenAI and Google bombed the field late at night, and Byte knocked down the price of the large model, Tencent couldn't wait to show its muscles.
On May 17, at the Tencent Cloud Generative AI Industry Application Summit, Tencent announced a series of product development progress:
Tencent's hybrid model capabilities continue to be upgraded, and the three versions will be fully open to industrial customers and individual developers.
Tencent's one-stop AI agent creation and distribution platform "Tencent Yuanqi" was officially launched, and the application experience is open from now on.
Tencent Yuanbao, an assistant app for individuals of Tencent's hybrid model, will be released on May 30;
Tencent Cloud has launched a large-model native toolchain, with three engine tools to help enterprises efficiently develop AI applications.
In addition, Tencent has also released a number of out-of-the-box SaaS products for enterprise customers and developers, and Tencent Cloud's AI development and engineering capabilities, such as the TI platform and vector database, have also been comprehensively upgraded.
Tencent ingots + yuanqi,
AI search and agents are all won
Previously, Tencent's hybrid assistant could only be used on WeChat mini programs or websites (https://hunyuan.tencent.com/bot), and soon the mobile app will be launched.
At this conference, Tencent announced that it will launch the assistant app "Tencent Yuanbao" for individuals on May 30.
According to reports, this is an efficient information integration tool driven by a large model and a search engine, with a simple interface design, which can search for real-time information, summarize and translate uploaded multi-format documents, and practice oral language with voice dialogue.
It seems that the AI search war is about to usher in a blockbuster, and the AI assistant App supported by Tencent's content ecology is performing. "Number One AI Player" will follow up with the evaluation, welcome to pay attention.
In addition to Tencent Yuanbao, Tencent also released another product with a similar name, "Tencent Yuanqi".
Tencent Meta is a one-stop AI agent creation and distribution open platform based on the hybrid model, which allows users to create exclusive AI agents with a low threshold through prompts, plug-ins, workflows, AI-assisted creation and other capabilities.
At the same time, Tencent Meta pre-integrates Tencent's ecological characteristic plug-ins and knowledge base resources, and will also open third-party capabilities to provide creators with out-of-the-box tools.
The created agents can also be published to Tencent's global distribution channels such as QQ, WeChat, and Tencent Cloud, and high-quality agents have the opportunity to receive traffic support.
It can be seen that the market competition for AI agents is fierce, and intelligent twins platforms such as Tencent Yuanqi, Byte's button, and OpenAI's GPTs are competing for developers and users, hoping to build a prosperous ecosystem based on their own large models.
From now on, Tencent Yuanqi can apply for experience through the official website: https://open.hunyuan.tencent.com/my-creation
The "number one AI player" got the qualification for the internal test for the first time, and on the agent store page, there are 6 categories, including characters, efficiency tools, leisure and entertainment, game pastime, life assistant, and emotional helper.
For example, ask a "high EQ bot", how to politely refuse to work overtime on weekends with the boss?
Boss, I understand that sometimes the urgency of a project requires us to sacrifice our time off to work overtime. However, I also believe that work-life balance is very important for both productivity and personal health......
You can also ask the agent expert to analyze the name "Zhen Huan", judging from this answer, it is reasonable to suspect that Mixed Yuan has secretly passed the love pass 10 times.
The name "Zhen Huan" is like a beautiful picture, showing a female image that is both smart and kind. The owner of this name may have demonstrated extreme intelligence and emotional intelligence in life, being able to navigate complex interpersonal relationships with ease.
If you want to create an agent yourself, it's very easy to operate. Enter the name and introduction to generate an avatar by AI, and then use AI to generate detailed settings, including roles, skills, principles, etc., and enter questions for debugging on the preview page on the right.
For example, we created a "local love story bot", which basically grasped the essence of local flavor, and the literature and art were mixed with greasy.
On the release page, you can see that Tencent Meta supports pushing to QQ, WeChat customer service, mini programs, and official accounts, but it needs to be reviewed by the platform before it can be opened to everyone.
The hybrid model is fully upgraded,
Video generation can be up to 16 seconds
Since its debut in September 2023, Tencent's self-developed hybrid model has continued to be iteratively upgraded, taking the lead in adopting a hybrid expert model (MoE) structure, which has promoted performance improvement and inference cost reduction. According to reports, Mixed Yuan has excellent performance in Chinese, especially in text generation, mathematical logic and multi-round dialogue.
In terms of text generation, Mixin improves the performance of answering "up-to-date" questions. For example, in the question "When was the last time the Qinling brown giant panda was photographed", Mixed Yuan can accurately answer the latest time, while ChatGPT can only answer based on news information.
In terms of mathematics and reasoning ability, the mixed element has also been greatly improved. For example, the four numbers 1, 4, 5, and 6 are counted as 24 points, and ChatGPT miscalculates them with basic subtraction and multiplication, while the mixed element uses a combination of subtraction, division, and parentheses to calculate 24.
At present, the hot multimodal field is also a direction that Tencent Hybrid is actively exploring.
A few days before the conference, Tencent just announced that the hybrid Wensheng graph large model has been fully upgraded and open sourced, which adopts the same DiT architecture as Sora, which can not only support Wensheng graphs, but also serve as the basis for multimodal visual generation such as video.
Lu Qinglin, head of Tencent's Wensheng Diagram, said in an interview with the media today that Wensheng Diagram has better scalability after changing from Stable Diffusion to DiT, and its semantic expression ability will be stronger and can accommodate more information.
ChatGPT supports further modification of images through text commands, and Hybrid can now make partial modifications to the generated images through multiple rounds of dialogue, such as putting a red hat on the rabbit in the picture, and its subject and background can remain basically the same.
At the same time, the hybrid element Wensheng diagram supports the generation of high-quality 2D and 3D game images, and can also efficiently synthesize product materials, which has been put into production in advertising scenarios, which can save production costs for customers.
It only takes 30 seconds to generate a 3D model from a single image, and it has been used in many industries such as animation, automobiles, and architecture.
According to Jiang Jie, vice president of Tencent and head of Tencent's hybrid model, the resolution and motion amplitude of the video generated by the hybrid element are better than those of Pika and Runway.
In March this year, Tencent Mix, Tsinghua University and the Hong Kong University of Science and Technology jointly launched the "Follow-Your-Click" video model, bringing a new way of interaction. Based on the input model, you only need to click on the corresponding area and add a small number of prompt words to make the originally static area of the picture move, and convert it into a video with one click.
Hyelement's next-generation video model is being fully upgraded based on ST-DiT, which can currently generate up to 16 seconds of video, and Q3 is expected to reach 30 seconds.
Lu Qinglin mentioned that the biggest technical difficulty now is that different modalities are not aligned, just like video and audio will be confused if you want to use a model to generate at the same time, and Tencent's hybrid team is solving the problem of multimodal alignment.
Open to developers in three versions,
The Wenshengwen model will soon be open-sourced
Openness is one of the key words of this conference, and Tencent Cloud is expanding the generative AI industry ecosystem around the hybrid model.
According to the latest April 2024 Report on Chinese Large Model Benchmark Evaluation, SuperCLUE, Tencent's mixed model ranks in the first echelon of domestic large models and is in the excellent leader quadrant.
At present, Hybrid has expanded to a trillion-level parameter scale, and in terms of basic large-scale model capabilities, Tencent Hybrid has launched three flexible model scale versions, namely Hunyuan-Pro, Hunyuan-Standard, and Hunyuan-Lite, which are fully open to enterprise and individual developers through Tencent Cloud.
Depending on your business needs, you can choose the most suitable model. Among them, the largest hybrid element hunyuan-pro is the best version of the current hybrid model, with a context length of up to 32k tokens, which can be used in complex text creation, logical reasoning, mathematical calculation, intelligent NPC, professional fields and other application scenarios.
Wu Yunsheng, vice president of Tencent Cloud and head of Tencent Cloud Intelligence, said that Tencent pays more attention to the improvement of large model capabilities and is committed to providing products with both capabilities and prices for the industry.
At this conference, Tencent also announced that it will open source the core Wensheng model of hybrid yuan in the third quarter of this year.
Tencent is training three sizes (S, M, L) models for open source use, and the S version is mainly deployed on mobile phones with 3B parameters; The M version is mainly deployed on the PC side, with 5B parameters. Version L is mainly deployed in the cloud/data center and is an MOE architecture model with 30B parameters.
In addition, Tencent Cloud has launched three PaaS products: "Large Model Knowledge Engine", "Large Model Image Creation Engine", and "Large Model Video Creation Engine", aiming to simplify the application of large models and promote the popularization of industrial AI.
The large model knowledge engine upgrades the large model knowledge question answering capabilities, simplifies the process of data engineering, model fine-tuning, and application development, and provides document parsing, vector retrieval, and multi-round rewriting to improve the accuracy and query efficiency of enterprise knowledge question answering. At present, Tencent Cloud's large-scale knowledge engine has been applied in various industries such as government affairs, finance, and education.
The image creation engine provides API technical services with AI image generation and editing capabilities, which can intelligently create image content based on text or images, including image stylization, AI photography, product background generation, and line drawing to meet the needs of different industry scenarios.
The video creation engine provides video generation and processing capabilities, and supports video translation, stylization, image dancing, and face fusion, and is suitable for scenarios such as short video platforms and film and television production.
At present, the hybrid model has been applied in more than 600 business scenarios of Tencent, and has been continuously optimized through the company's ecosystem.
For C-end users, such as WeChat Reading, new functions such as AI book asking and AI outline have been launched based on the hybrid model, which has greatly improved the user's reading efficiency and experience. Tencent's customer service has significantly improved the accuracy and efficiency of intelligent dialogue through model upgrades, and handles 1.5 million user requests per day.
For the B-side, such as Tencent Advertising's AI creative platform "Miaosi", it has improved advertising efficiency with the help of a large hybrid model, and Tencent Meeting's AI assistant has also improved meeting efficiency through model optimization, and its usage has increased by 20 times in four months.
"The creation of large models is only the starting point, and the goal is to implement technology into industrial scenarios and create value." Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of Cloud and Smart Industry Business Group, said.