SenseTime Group Announces 2023 Interim Results

Exponential Growth in Generative AI Related Business whilst

Upgrading the Full-line Business Through Large Models

[Hong Kong, August 28, 2023] SenseTime Group Inc. (“SenseTime Group” or the “Company”) today announced its interim results for the six months ended June 30, 2023 (“1H”).

In the 1H 2023, the Company achieved steady revenue growth. Leveraging its SenseCore AI infrastructure, SenseTime Group accelerated the research and development (R&D) of large language models (LLMs) and text-to-image models under the “SenseNova” foundation model sets, significantly broadening its Generative AI business. During the period, in collaboration with a number of top domestic research institutions, the Company released the pretrained LLM - InternLM, becoming the first domestic base model to surpass the performance of GPT-3.5-turbo. The Company also launched the Generative AI product suite on the basis of the “SenseNova” foundation model sets, including “SenseChat”, “SenseMirage”, “SenseAvatar”, and more. These innovations drove triple-digit growth in the Generative AI related business.

Dr. Xu Li, Executive Chairman of the Board and Chief Executive Officer of SenseTime Group, commented, “In 1H, large models and Generative AI made some of the world's most eye-catching technological breakthroughs. As such, SenseTime is going through an extremely critical period of development. Our ‘SenseNova’ foundation model sets and the SenseCore AI infrastructure are performing impressively. We are seizing new opportunities fueled by technological breakthroughs and successfully released the first domestic large model to surpass GPT-3.5-turbo, demonstrating comprehensive capabilities. Subsequently, revenue from Generative AI related business surged 670.4%. We look to expand our large model capabilities within the industry to help our users to successfully make disruptive products in the era of Generative AI.”

Key Highlights of the Company’s Business Operations in the 1H 2023:

● The Company introduced the “SenseNova” foundation model sets and released InternLM, the first domestic base model to surpass the performance of GPT-3.5-turbo.

● The text-to-image product SenseMirage was released and upgraded to version 3.0. Its image generation capabilities rank among the global leaders and hold a leading position domestically.

● Continued to upgrade the SenseCore AI infrastructure, with the number of GPUs increasing from 27,000 at the end of March 2023 to approximately 30,000 at present, resulting in an improved computational capacity by 20% to 6 ExaFLOPS.

● Various business sectors have been fully empowered by large models and Generative AI capabilities. This has resulted in a remarkable 670.4% growth in Generative AI related revenue, infusing the company with renewed vitality and momentum.

● Revenue performance was stable overall, with RMB1.43 billion in total revenue, representing a year-on-year increase of 1.3%; the gross profit margin was 45.3%.

● Net loss was reduced by 2.0% year-on-year to RMB3.14 billion and the adjusted net loss reduced by 6.7% year-on-year to RMB2.39 billion.

● Total cash reserves amounted to RMB14.8 billion.

LLMs

The Company collaborated with multiple top domestic research institutions, investing around 10,000 GPUs in the R&D of LLMs, enabling a significant leap in model performance:

● In April, the Company introduced “SenseChat 1.0”, one of China’s earliest chatbot products based on an LLM with hundreds of billions of parameters.

● In June, the Company released the InternLM with 104 billion parameters, which is the first base model with comprehensive capabilities to surpass GPT-3.5-turbo. It was trained on a multilingual corpus of 1.6 trillion tokens, supporting over 20 languages. “SenseChat 2.0” was launched in early July based on this base model.

● In August, the Company finished training InternLM-123B, an upgraded version with 123 billion parameters. InternLM-123B ranks second in the world for the text on a total of 300,000 questions in 51 well-known evaluation benchmarks (including MMLU, AGIEval, ARC, C-Eval, RACE, GSM8K and etc.), outperforming GPT-3.5-turbo and Meta's newly released LLaMA-2-70B. SenseChat will be upgraded to version 3.0 based on InternLM-123B in September.

● SenseTime is ahead of the industry in terms of the accumulation of raw corpus data and the ability to conduct cleaning and testing of corpus. The Company currently has achieved a monthly output of around 2 trillion high-quality tokens and expects the high-quality data storage to surpass 10 trillion tokens by the end of the year, supporting the training of even more powerful base models.

● The Company supports the development of the open-source platform. Parts of InternLM-7B’s (7 billion parameters) training data, training code, and weights have been opened to academia and industry at no cost, and supports commercial use. InternLM-70B ranks the highest in multiple model benchmarks for lightweight base models.

Text-to-Image Model

The Company released a leading text-to-image model in China. To this end, approximately 2,000 GPUs were allocated in the development of the text-to-image model, achieving rapid iterations:

● In January, the SenseMirage 1.0, powered by self-developed 800 million parameter diffusion model by SenseTime, began internal testing. Users were able to input prompts to generate anime and manga images.

● In April, SenseMirage was upgraded to version 2.0 and was officially released as SenseTime’s first for-consumer application based on a 1.4-billion parameter text-to-image base model with general capabilities.

● In July, SenseMirage was upgraded to version 3.0, with base model parameters increasing to 7 billion. Supported by a new, large, self-developed generative model architecture, it provides stronger image generation effects and details, achieving professional photography-level image depiction. The core algorithm of SenseMirage 3.0 surpasses both Google’s Imagen and OpenAI’s DALL·E 2 on the COCO benchmark. Its image generation capabilities rank among the global leaders and holds a leading position domestically.

● SenseMirage 4.0 is expected to be launched in Q3 2023. Along with the upgraded base model, it will also introduce an entirely new aesthetic system to further enhance the product’s competitiveness.

Business Review Highlights

Comprehensively Promote the Generative AI Business and Empower Industry Innovation

● Based on the “SenseNova” foundation model sets and SenseCore, the Company launched the Generative AI product suite, consisting of “SenseChat”, “SenseMirage”, “SenseAvatar”, “SenseSpace”, “SenseThings”, and “SenseAnnotation”. These products cater to five mainstream Generative AI applications, including natural language interaction, AI-assisted image and text generation, digital avatars, 3D large-scale scene reconstruction, and 3D object generation. These new capabilities were promptly provided to customers across the Company's four major business segments: Smart Business, Smart Life, Smart Auto, and Smart City, enabling customers to quickly iterate on new products leveraging large models.

● In the first half of 2023, the Generative AI related business recorded a year-on-year revenue growth of 670.4%, and its contribution to the Company’s business rapidly increased to 20.3%. According to the “AI Large Model Market Research Report (2023)” by Frost & Sullivan, the Company ranks first in China for overall competitiveness, including product, strategic vision, and ecosystem development.

● Key highlights of the Company’s four business segments in the 1H 2023:

Smart Business: The Company has collaborated with hundreds of clients across ten different industries to customize large models. The average revenue per customer increased by 59.9%.

Smart Life: The customers in the smartphone and mobile internet segment demonstrated active involvement in innovative initiatives centered around Generative AI. As a result, the customer base experienced a rebound and significant growth.

Smart Auto: Shipment-based pre-installed productions business achieved a year-on-year growth of 573%, with the number of units delivered reaching 390,000. The production also led to an increase in gross profit per vehicle of approximately 29%. While research and development (R&D) revenue declined, there was a rapid increase in the proportion of revenue generated from mass production.

Smart City: The Company continued to advance the transformation of the business, with a focus on serving premium clients and improving cash flow.

● On the international front, the Company has been consistently expanding its overseas business. Its revenue from overseas market for the 1H was RMB229.2 million and the percentage of revenue contribution is 16.0%, comparable to the full year of 2022. The Company’s proprietary LLM, together with its capabilities to adapt to different languages and cultures around the world, are favored by many overseas customers, making its global partners true AGI beneficiaries.

Environmental, Social and Governance

SenseTime has always been committed to building a trustworthy AI system and actively participated in the formulation of domestic AI standards and regulations, so as to enhance AI safety and governance and promote sustainable AI development. During the period, the Company has launched SenseTrust, an AI governance platform which encompasses a comprehensive suite of trustworthy AI governance tools for AI training data, model and applications.

At the same time, the Company continues to strive for carbon neutrality goals of “Carbon Peak by 2025, Carbon Neutrality in Operations by 2030, Net Zero Emissions by 2050”.