occurrences

On July 4-6, during the 2024 WAIC World Artificial Intelligence Conference, we attended the “Great Love Without Frontiers · Xiang Xinli” AI Forum hosted by Shangtang Technology. On July 5, at the forum, Shang Tang unveiled the first large-scale streaming multi-modal AI interaction model “Nishi-Nissin SenseNova 5.5,” and launched the upgraded end-side model “Nishi-Nissin 5.5 Lite”. At the same time, he shared technical experience and progress in the fields of big models, computing power infrastructure, intelligent driving, etc., and showcased practical results empowering many industries. The overall performance of “Nishi-Nisshin 5.5” has been increased by 30% compared to “Nishi-Nisshin 5.0”, and its interactive effects and many core indicators can be compared with GPT-4o.

The accuracy of the “Nichi-Nisshin 5.5 Lite” has been improved by 10% compared to the 5.0 version, the inference speed is 15% faster, and the first package delay has been reduced by 40%.

reviews

Combining many innovations, the “Japanese New 5.5” performance can be compared to GPT-4O: 1) integrates cross-modal information to provide users with real-time streaming multi-modal interaction, innovates the AI interaction mode, achieves “what you see is what you get”, and is also more suited to simultaneous processing of multiple tasks such as scene understanding and overview; 2) Shang Tang pioneered the construction of a thought chain and verified the correctness of each step of the thought chain with advanced technology, thus achieving large-scale production of high-quality thinking chains, strongly enhancing the reasoning and thinking ability of large models, and clearly following improvements in mathematical logic, English, and instructions..

Strong accumulation, low cost+large-scale capability to support efficient iteration of technology, shortening the gap between the Chinese and US SOTA models: OpenAI released GPT-4O on May 13 this year, and it has been more than half a year since GPT-4 Turbo was launched on November 6, 2023; however, after announcing “Rijixin 5.0” on April 23 this year, Shang Tang successfully iterated “Ririxin 5.5” in just over 50 days, making up for the shortcomings of the previous generation's capabilities, and the multi-modal capabilities were significantly improved. It can be compared to GPT-4o, which greatly shortens the cycle of catching up with the target model, and highlights its industry-leading R&D capabilities. The model has helped expand applications in vertical industries such as the Internet, finance, and healthcare, and is now being used by more than 3,000 domestic service enterprise customers.

At present, the technical verification of “Nishi-Nisshin 6.0” has also been completed, and it is at the top of all models below 12B in the HuggingFace global ranking. In the second half of the year, the company plans to stop distinguishing between language models and multi-modal models, but instead position the final form of version 6.0 as an integrated model, continue to optimize streaming interaction, expand the boundaries of capabilities, cover a wider range of application scenarios through distributed computing with endside cloud collaboration, and achieve an order of magnitude reduction in inference costs to support sustainable user volume. In the future, Shang Tang will also rely on its deep technical heritage and experience to incorporate more modes such as three-dimensional spatial information into multi-modal capacity building.

Continue to cultivate the end side and promote deployment penetration: The upgraded end-side model “Nichi-Nisshin 5.5 Lite” is at the leading level in the industry in all fields of end-side applications, provides high-speed inference, and has a very low initial delay, further optimizing the user experience. At present, the “Nishi-Nissin” end-side model has penetrated into various industries and is commercially connected with more than 150 customers, covering IoT device deployment applications such as smartphones, tablets, VR all-in-one machines, in-vehicle computers, smart desk lamps, etc.; it has been integrated into the “Daily Sun? The cost of a single device for the large “discussion” end side model is as low as 9.9 yuan per year.

Shang Tang Yueying was the first in the industry to implement a native multi-modal large model for vehicle side deployment. It can be deployed using full-stack methods such as cloud side, cloud side integration, and end side. The vehicle-side 8B multi-modal model reduced the first packet delay to less than 300 ms, and the inference speed reached 40 tokens per second. Shang Tang Jueying has now released the industry's first generative interactive interface products, “Anytime Interface” and “Arbitrary Control”, and other in-vehicle AI agents, to enrich the supply of automotive intelligent products and promote the deep integration of multi-modal large models with smart cars.

The generative AI business is growing rapidly, and profits can be expected at the group level: continuous advancements in technology are driving strong growth in the company's generative AI business, which accounts for a significant increase in revenue share within the group, and is expected to account for more than 50% this year. Shangtang has a leading computing power base in China and serves a large number of customers with strong computing power requirements. Recently, Shangtang Dajiao Computing Power Service passed the national computing power service capability maturity enhancement level assessment, becoming the first company to pass this assessment.

At this stage, the company uses profits generated in the field of computing power to subsidize large-scale model research and development, and is actively investing in generative AI business. At the same time, it is actively connecting financing and introducing incremental capital for some non-core business sectors to meet their high capital requirements in the short to medium term, with the goal of achieving profit at the group level in the next 1-2 years.

Investment advice: We maintain our previous target price of HK$1.5 and a better rating than the market.

risks

Industry competition intensifies