Huafu Securities: Operators' Wanka Cluster Plans Are Surfacing One After Another, Ola Friend Drives Popularity of the Doubao App

Zhitongcaijing · 10/17 06:09

The Zhitong Finance App learned that on October 12, Huafu Securities published a research report saying that on October 12, at the 12th China Mobile (00941) Global Partner Conference, China Mobile stated that it will vigorously implement the “AI+” action plan, which includes building an intelligent computing cluster of over 10,000 cards, depositing trillions of tokens industry data sets, and training a large AI model with trillion parameters. The progress of big models reflects the rapid development of Chinese enterprises in the field of AI, especially breakthroughs in big model training and application, and will also provide stronger technical support for the intelligent transformation of various industries. The Wensheng video AI model at home and abroad continues to be updated, and it will challenge OpenAI and Meta (META.US). OlaFriend, the first AI smart headset released by Byte, was connected to the big Doubao model, driving the number of downloads of the Doubao app. On October 11, the Doubao app had over 4 million downloads in a single day.

1. Operators' 10,000 card cluster plans have surfaced one after another, empowering thousands of industries to continue to be carried out

China Mobile is advancing the “AI+” program to build an intelligent computing cluster. On October 12, at the 12th China Mobile Global Partner Conference, China Mobile stated that it will vigorously implement the “AI+” action plan, which includes building a super10,000 card intelligent computing cluster, depositing trillions of tokens industry data sets, and training a large AI model with trillions of parameters. (1) In terms of intelligent computing, China Mobile's two major intelligent computing centers with 19,000 and 18,000 GPU cards, respectively, in Hohhot and Harbin, have been put into operation, and are the largest single intelligent computing centers for global operators. (2) In terms of large model applications, China Mobile released the Jiutian Zhizhi multi-modal base large model and launched 30 self-developed industry models. The model is a full-stack localized full-modal base model based on Wanka's domestic computing power cluster and domestic algorithm framework. It has been significantly improved in four major functions: intelligent analysis of long texts, full-duplex voice interaction, video and image processing, and deep insight into structured data.

CNOOC released the “Haineng” artificial intelligence model. On October 14, CNOOC collaborated with companies such as China Telecom and iFLYTEK to launch the “Haineng” artificial intelligence model. The model relies on massive data resources and integrates existing models, taking into account both professional and general scenarios. At this press conference, 5 professional scene models and 6 general scene models with HNA characteristics were specially launched.

Earlier, Huafu Securities mentioned that China Telecom completed the first domestic model with large trillion parameters based on the training of a nationally produced 10,000 card cluster, and improved the large star semantic model TelecHat2-115b. Overall, these developments reflect the rapid development of Chinese enterprises in the field of AI, especially breakthroughs in large-scale model training and application, and will also provide stronger technical support for the intelligent transformation of various industries.

2. The Wensheng video AI model at home and abroad continues to be updated, and it will challenge OpenAI and Meta

On October 14, Adobe announced a public test of the Wensheng video AI model Firefly. Adobe said it has made a new breakthrough in its technology and will compete with OpenAI and Meta. Unlike its competitors, Adobe claims that it uses exclusively licensed content for training, which is expected to circumvent the ethical and copyright issues faced by other generative AI tools, and is “the first publicly available commercial security video model.” However, Adobe has yet to announce its official launch date and will only provide access to users on the waiting list during the testing phase. On October 12, Kuaishou jointly launched the PyramidFlow open source Wensheng video AI model. After entering text, users can generate high-quality video with a maximum length of 10 seconds, a resolution of up to 1280x768, and a frame rate of 24 fps. The model is characterized by efficient generation, fast reasoning, and open source and commercial use.

3. The smart spectrum open source Wensheng chart model is online, and the overall performance is outstanding

On October 14, the Smart Spectrum Technology Team released the open source Wensheng Chart model CogView3 and

CogView3-plus-3b, the capabilities of this series of models have been launched on the “Smart Spectrum Clear Statement” app. CogView3 is a Text2IMG model based on cascaded diffusion. In terms of actual effect, CogView3 is 77.0% higher than SDXL, the most advanced open source text-to-image diffusion model SDXL, and only requires about 1/10 of SDXL's inference time. The CogView3-Plus model introduced the latest DIT framework based on CogView3 and introduced a text-image joint attention mechanism to further improve overall performance.

4. AI big model tracking

Byte OlaFriend connects to the big Doubao model, driving the number of downloads of the Doubao app. According to SimilarWeb's latest data (1004-1010): (1) Pika's visits continued to rise this week. With the release of Pika 1.5, the popularity of Pikaffects' special effects continued to rise during this week. (2) Following a decline in the previous two weeks, the number of visits to Runway returned to an upward trend this week. On October 9, Runway released an update to support first and last frame control. According to the latest data from Qimai (1009-1015): (1) Kimi's download volume has increased. On October 11, the Kimi Discovery Edition, which has AI independent search capabilities, was launched. The search volume is 10 times that of the regular version, and you can read 500 pages in one search. (2) On October 11, the Doubao app had more than 4 million downloads in a single day. On October 10, ByteDance Doubao released the first AI smart headset OlaFriend. It is connected to the Doubao Big Model and is deeply integrated with the Doubao App. After users put on the headset, they can invoke Doubao through voice without turning on the phone.

Recommended to follow

Domestic computing power chains: Cambrian (688256.SH), Haiguang Information (688041.SH), Longxin Zhongke (688047.SH), Shengke Communications (300685.SZ/00829).

Data center hardware: Zhongji Xuchuang (300308.SZ), IFF (601138.SH), Xinyisheng (300502.SZ), Tianfu Communications (300394.SZ), Wall Nuclear Materials (002130.SZ), Invec (002837.SZ), Shanghai Electric Power Co., Ltd. (002463.SZ), Shenghong Technology (300476.SZ).

Risk Alerts

Technology development and implementation fell short of expectations; downstream terminal shipments fell short of expectations; downstream demand fell short of expectations; market competition increased risks; geopolitical risks.