IFlytek executives interpret financial reports: The Spark model will achieve benchmarking, ChatGPT will not be affected by Nvidia

Lei Diwang Lotte October 20th

IFLYTEK released its financial report yesterday. According to the financial report, iFlytek's revenue for the first three quarters of 2023 was 12.614 billion yuan, a decrease of 0.37% compared to the same period last year; The net profit was 99.36 million yuan, a decrease of 76.36% compared to the same period last year; The net loss after deducting non income was 324 million yuan, and the net profit after deducting non income in the same period last year was positive.

IFlytek's revenue in the third quarter of 2023 was 4.772 billion yuan, an increase of 2.89% compared to the same period last year; The net profit was 25.79 million yuan, a decrease of 81.86% compared to the same period last year; The net loss after deducting non income was 20.19 million yuan, and the net profit after deducting non income in the same period last year was positive.

After the release of the financial report, Jiang Tao, the Secretary of the Board of Directors and Vice President, communicated with investors. Jiang Tao revealed on a conference call that on October 24th, the Spark Cognition Big Model V3.0 will be released as scheduled, and the seven dimensions of capabilities will be comprehensively improved. The code capabilities of the big model will be upgraded to achieve comprehensive benchmarking with ChatGPT.

The iFlytek Spark model has been deeply applied in various business fields, continuously improving the quality of iFlytek's products and services in markets such as education, healthcare, smart cities, operators, finance, and automobiles, enhancing user experience, and reducing research and development and delivery costs

On October 17th, the US government further tightened export controls on cutting-edge artificial intelligence chips, strictly restricting Nvidia and other chip manufacturers from selling high-performance semiconductors to China. According to the new regulations, regulated Nvidia chips involve A100, A800, H100, H800, L40, and even RTX4090, which will have a significant impact on large model manufacturers who build large models on Nvidia's computing power.

Regarding this, Jiang Tao pointed out that the iFlytek Spark model is not affected by Nvidia because iFlytek has already been sanctioned by the United States, and the current domestic large model architecture is based on independently innovated software and hardware. The premise of safety and controllability is that computing power is controllable. Currently, the Huawei Shengteng 910B has basically achieved the ability to benchmark against the NVIDIA A100.

The following is the Q&A session for iFlytek executives:

Q: What will be the impact of further restrictions on NVIDIA's export of high-end chips to China by the United States on the development of artificial intelligence in China?

1On October 17th, the US government further tightened export controls on cutting-edge artificial intelligence chips, strictly restricting Nvidia and other chip manufacturers from selling high-performance semiconductors to China. According to the new regulations, regulated Nvidia chips involve A100, A800, H100, H800, L40, and even RTX4090, which will have a significant impact on large model manufacturers who build large models on Nvidia's computing power.

2. After being included in the US entity list in 2019, iFlytek was once again sanctioned by the US on October 7, 2022, against 28 leading Chinese companies and institutions in the fields of artificial intelligence, high-performance chips, and supercomputers, including iFlytek.

In the context of backward computing power at that time, iFLYTEK continued to improve its algorithms. For example, based on iFLYTEK's leadership in undertaking many projects on the national autonomous and controllable artificial intelligence platform in the past few years, it achieved nearly a thousand times the acceleration of the inference efficiency of the multi billion parameter large model in engineering technology, providing the possibility for the large-scale application of larger and more cognitive intelligence large model technologies.

At the beginning of 2023, iFlytek and Huawei Shengteng launched a special research and development project, jointly developing a high-performance operator library with Huawei, and working together to create a new base for general artificial intelligence in China, allowing domestic large models to be built on the basis of independently innovative software and hardware.

3. The key to the deep application of cognitive big models in the industry is safety and controllability, scenario driven, and exclusive controllability. The premise of safety and controllability is computing power controllability. Currently, the Huawei Zengteng 910B capability has basically achieved benchmarking with the NVIDIA A100. At the upcoming iFlytek 1024 Global Developers' Festival, iFlytek and Huawei will further jointly release their AI computing power base.

4. It is precisely because iFlytek's Spark model has achieved autonomy, controllability, and technological self-reliance that the Spark model has the opportunity to go abroad and provide AI services to billions of users in different countries and languages around the world in the future.

Q: May I ask what is the overall progress and pace of implementation of the ToB market for the current Spark model for industry and corporate customers?

Answer: 1. Products aimed at enterprises and industry customers include common products that are ready to use out of the box, such as code, customer service, office, etc., as well as specific applications related to industry knowledge such as marketing and enterprise knowledge base. Based on the general model, specialized training and tuning are conducted for specific fields. In terms of product application effectiveness, taking the code product as an example, the iFlyCode1.0 programming assistant of iFlySpark can automatically generate and modify according to needs. Based on the use of iFlySpark's internal R&D personnel in tool development, script writing, data analysis, and other scenarios, the overall coding efficiency has increased by more than 50%.

2. There are two main ways to deliver to B-end customers:

Method 1: Provide enterprise services through cloud services, which is currently adopted by OpenAI. On August 29th, OpenAI launched ChatGPT Enterprise, which allows customers to input company data and train and customize ChatGPT for their industry and use cases.

Method 2: To deliver privatization solutions to high-end customers who emphasize data security, there are generally two delivery modes: first, to deliver domestically produced and controllable software and hardware integrated spark machines. The iFlytek Spark all-in-one machine, jointly launched by iFlytek and Huawei, has a performance that can benchmark against the A100. It can customize exclusive large models for each industry partner and leading enterprise, including scenario development platforms and rich scenario packages. Secondly, based on the customer's own hardware, iFlytek provides a "software+service" solution for the privatization deployment and training of the Spark Big Model.

3. In terms of landing rhythm, combined with demonstration projects from benchmark customers, iFlytek Spark will gradually launch privatization solutions in fields such as operators, finance, automobiles, and industry.

The overall process mainly includes: technical research, product selection, project initiation, feasibility study, network connection, bidding, project construction, and acceptance. Customers generally need at least 6 months from technical research to bidding. Currently, privatization deployment is in the stage of scheme docking, effectiveness verification, and testing. Xunfei will jointly develop relevant product services with customers to help them improve business efficiency and product competitiveness, and participate in sharing and creating value together.

Q: What is the overall situation of the Starfire model empowering the company's existing business and products?

Answer: The iFlytek Spark model has been deeply applied in various business fields, continuously improving the product service quality of iFlytek in markets such as education, healthcare, smart cities, operators, finance, and automobiles, enhancing user experience, reducing research and development and delivery costs, and enhancing the market competitiveness and user penetration rate of existing products.

Taking education as an example, on August 15th, at the press conference of iFlytek's Spark Model V2.0, a smart classroom equipped with the Spark Cognitive Model was upgraded and the Spark Teacher Assistant was launched. Through dialogue and generative interaction, the teaching design of the intelligent generation science system for teachers was designed to improve the efficiency of unit teaching and lesson preparation, continuously enhancing the competitive advantage of business and products.

Since the official release of the Spark Model on May 6th, it has brought continuous promotion to the sales of the company's learning and office hardware products. Taking the learning machine as an example, with the support of Spark, the iFlytek AI learning machines T20 series and Lumie10 series have created a complete set of exclusive scientific learning solutions for children through personalized precision learning systems and seven one-on-one, one-on-one, and one-on-one human assisted learning functions, including English speaking practice, Chinese English essay correction, interactive learning in mathematics, encyclopedia free Q&A, parent-child education assistant, intelligent programming assistant, and creative painting partner.

Q: What is the usage and overall feedback of the large model capability call for developers?

Answer: Since the release of the Spark Model, the iFlytek Artificial Intelligence Open Platform has added more than 1.4 million developers, reaching 5.49 million in October. It has opened up a total of 595 AI capabilities, and developers have developed over 12000 assistant applications. Currently, it has cooperated with over 10000 enterprises for docking.

On September 5th, the iFlytek Spark API was fully opened to developers, offering multiple package types for two different versions, V1.5 and V2.0.

V1.5 supports multiple abilities such as language comprehension, knowledge Q&A, code writing, logical reasoning, and mathematical problem-solving, with built-in online search capabilities. This version currently offers four packages: Package 1 (100 million tokens, 0.14 yuan/10000 tokens), Package 2 (50 million tokens, 0.15 yuan/10000 tokens), Package 3 (5 million tokens, enterprise free package), and Package 4 (2 million tokens, personal free package);

V2.0 is comprehensively upgraded on the basis of V1.5, and has been specially upgraded in code and mathematical scenarios. The service has embedded functions such as online search, date query, weather query, stock query, poetry query, and word understanding.

Developers with different usage needs can choose different package services. This version currently offers four packages: Package 1 (100 million tokens, 0.28 yuan/10000 tokens), Package 2 (50 million tokens, 0.3 yuan/10000 tokens), Package 3 (enterprise free package, 5 million tokens), and Package 4 (personal free package, 2 million tokens).

Q: Can I share the relevant usage data indicators of the Spark Model APP?

Answer: 1. The number of registered users, traffic volume, daily activity, monthly activity, etc. of iFlytek Spark are all in the first tier of domestic large-scale models. On September 5th, iFLYTEK announced that the iFLYTEK Spark cognitive model was open to the public. Within 14 hours of its launch, iFLYTEK Spark had over 1 million users and quickly ranked first on the App Store free total ranking.

2. The peak usage period of iFlytek Spark users is concentrated in the two periods of "10 o'clock and 15 o'clock". Compared with other internet manufacturers' large models that focus on entertainment products, the iFlytek Spark model emphasizes "liberating productivity and unleashing imagination". The peak usage period of the above users also fully verifies the usage habits of iFlytek Spark that are closer to the office population.

3. On June 9th, Spark Assistant was released. Spark Assistant is an efficient productivity tool based on the iFlytek Spark model, designed for user usage scenarios, and can quickly meet user task needs in various scenarios. As of now, developers have developed over 12000 assistants using the Spark app, covering marketing, creative, programming, chat, learning, reasoning and other scenarios.

4. IFlytek insists on continuously improving the user value and experience of the Spark app, solidly developing users. We will not develop users through burning money, but as the product user value increases and reputation spreads, we will definitely attract more and more users.

Question: What are the highlights of the 1024 press conference?

Answer: In this 1024 Developer's Day, the Spark Cognition Model V3.0 will be released as scheduled, with comprehensive improvements in the seven dimensions and upgraded code capabilities for the big model, achieving comprehensive benchmarking of ChatGPT. At the same time, release the medical big model and Xiaoyi APP to provide health assistants for every family; Newly released AI Q&A tutoring and AI psychological counseling, providing growth partners for every child; For the first time, a large model of scientific and technological literature has been released to provide spark assistance for every scientific researcher; Collaborate with industry leaders to release 12 industry models and provide productivity tools; Joint release of the latest cooperation results with Huawei on the large model computing platform.

Q: What is the overall trend of education business in the first three quarters? What is the proportion of GBC?

1. The educational concept of personalized learning and individualized teaching by iFlytek has been recognized by the education regulatory department. On October 13th, the People's Daily elaborated on the promotion of digital education and AI empowerment education as a key direction that the country continues to pay attention to and guide.

2. The AI education products equipped with iFlytek Spark have created significant application value for teachers and students. The Spark Teacher Assistant, released on August 15th, uses dialogue and generative interaction to intelligently generate scientific system teaching designs, flexible and practical teaching activity designs, courseware tailored to teaching needs, and other content for teachers, improving the efficiency of unit teaching lesson preparation and assisting in new changes in classroom teaching.

With the support of Spark, the iFlytek AI learning machines T20 series and Lumie10 series have created a complete set of exclusive scientific learning solutions for children through personalized precision learning systems and seven one-on-one, one-on-one human assisted learning functions, including English speaking practice, Chinese English essay correction, mathematical interactive learning, Baike free Q&A, parent-child education assistant, intelligent programming assistant, and creative painting partner.

3. In terms of the overall development of education business, compared to the funding shortage during the unified construction process of regional level individualized education solutions with government coordinated funds, iFlytek's school-level individualized education solutions for schools use educational information funds within a guaranteed budget. With the advantages of iFlytek's core technology competitiveness and product application effects, the market share continues to increase and maintain a good growth trend.

4. The education business has overcome the difficulties faced by the G-end business and overall maintained healthy growth. The revenue in the first three quarters increased by 6.78% year-on-year, and the business has been improving quarter by quarter. The business structure has been continuously optimized: in 2022, the G-end accounted for 40%, the B-end accounted for 30%, and the C-end accounted for 30%; In the first three quarters of 2023, the G-end accounted for 25%, the B-end accounted for 32%, and the C-end accounted for 43%.

Lei Di was founded by media person Lei Jianping. If reprinted, please specify the source.