Xinghuo cognitive model version 2.0 failed to boost the stock price, what concerns do investors have about iFLYTEK?

Interface News Reporter | Yu Hao

Interface News Editor |

On August 15, iFLYTEK released the 2.0 version of the iFLYTEK Spark model as scheduled, which has upgraded its code capabilities and multi-modal capabilities compared with version 1.5.

Based on the above capabilities, iFlyCode, a programming assistant that supports rapid code generation and bug change, and a spark teacher assistant that generates courseware with one click. Benefiting from the improvement of multi-modal capabilities, the capabilities of applications such as Xinghuo Language Partner and iFLYTEK Smart Work based on the Spark large model have also been improved.

In addition to software products, iFLYTEK and Huawei also jointly released the iFLYTEK Spark all-in-one machine. According to Liu Qingfeng, chairman of iFLYTEK, the device can support the training, inference, customization and optimization of large models, and can benchmark NVIDIA's A100 chip in single-card computing power.

When releasing the Spark model on May 6, iFLYTEK gave a milestone plan for the release of V1.5 on June 9, V2.0 on August 15, and ChatGPT, Chinese surpass, and English equivalence on October 24. In the past, iFLYTEK's stock price rose sharply after the press conference, but this time, its stock price performance fell all the way to the limit.

As of the close of trading on August 16, iFLYTEK reported 58.48 yuan per share, down 8.63%.

The ambition of building an industry ecology is highlighted

Judging from the overall product information disclosed, the content of iFLYTEK's press conference is remarkable.

For large models, code capabilities are not only the embodiment of reasoning capabilities, but also reduce the development threshold of large model ecosystems, which is of great significance for expanding application ecosystems.

Liu Qingfeng pointed out at the press conference that iFLYTEK Spark 2.0 has upgraded code capabilities in five dimensions, including code generation, code completion, code error correction, code explanation, and unit test generation. According to the OpenAI-built code ability public test set HumanEval, the effect of the Spark V1.5 Python language is only 41 points, and V2.0 has reached 61 points, close to ChatGPT.

Based on code capabilities, iFLYTEK also brings the programming assistant iFlyCode 1.0 to developers. In the live demonstration session at the press conference, Liu Cong, Dean of iFLYTEK Research Institute, entered the iFlyCode interface to complete several steps of Prompt, and completed the programming of the sample application without writing code directly. iFLYTEK feedback said that in some typical scenarios, the iFlyCode code adoption rate reached 30%, the coding efficiency increased by 30%, and the overall efficiency increased by 15%.

In terms of multimodal capabilities, according to reports, the iteration of iFLYTEK Spark model is mainly reflected in image description, image question and answer, image recognition, text and image generation, virtual human synthesis, etc. At the press conference, the Spark model completed tasks such as "looking at pictures and speaking" and "painting according to ancient poems".

Focusing on multi-modal capabilities, iFLYTEK has also built related applications, through the video production assistant iFLYTEK Smart Work, users can complete the generation of virtual human short videos, copywriting, graphics, dubbing and other work can be completed with one click.

The above-mentioned ability of large model iteration is also applied to iFLYTEK AI learning machine, which adds AI programming space, AI creative drawing board, and AI speaking related functions. After the application of Xinghuo model, the sales of iFLYTEK AI learning machines increased significantly, and the proportion of C-end revenue in iFLYTEK's education revenue in the first half of the year has increased from 28% in 2022 to 40% in 2023.

It is worth mentioning that at this conference, iFLYTEK and Huawei also jointly released the iFLYTEK Spark all-in-one machine, which is the first demonstration of achievements since the two announced the joint construction of the computing power base in July this year. According to reports, the device can provide 5 customized optimization modes such as dialogue development, task orchestration, plug-in execution, knowledge access, prompt engineering, as well as more than 10 scenario packages such as office, code, customer service, operation and maintenance, marketing, procurement, etc., and support 3 model sizes for users to choose.

From the content released, it can be seen that iFLYTEK's ambition to build an industry ecology around iFLYTEK Spark. In December last year, iFLYTEK pointed out that "1" refers to the general cognitive intelligence model, and "N" refers to the landing of large models in education, office, automobiles, human-computer interaction and other fields.

Measures such as providing programming assistants for ecological developers and providing computing power bases for industry customers indicate that iFLYTEK hopes to become a platform and foundation for all walks of life to test the big model. In the group interview after the press conference, Liu Qingfeng once again emphasized that doing a good job of the general large model is the premise of doing a good job in the vertical large model, and made it clear that a more scientific way in the future will be for some sub-industry leaders who understand data and scenarios to cooperate with iFLYTEK, and iFLYTEK will provide a general capability base and training platform, and customers will optimize on this basis.

But he also said that the core technology expressed in the form of software in the market has not been reasonably priced, so in order to make a profit, iFLYTEK also needs to do some vertical products. "In this process, we will make it clear to the developer that if the developer competes with us, iFLYTEK will not carry out technical empowerment, so as to ensure that what he does and we will not crash." Liu Qingfeng added.

What are the market concerns about the decline in stock prices?

Since the release of the Spark Big Model, the news related to the big model can almost be regarded as the direction of iFLYTEK's market sentiment.

Looking back at its stock price trend, after the release of its first-quarter earnings report at the end of April, its stock price fell once due to the year-on-year decline in its net profit and other financial indicators; This trend was reversed after the release of the Spark model in early May, which rose by 10.01% in the first trading day after the release.

Affected by the topic of large models, as of the close of trading on August 16, iFLYTEK's share price has increased by 70.6% compared with the closing price on January 3 at the beginning of this year. After the release of the Spark Model V2.0, the stock price did not rise, but once fell to the limit. After-hours data released by the exchange showed that Guolian Securities Shanghai Feihong Road sold 159 million yuan, an institution sold 132 million yuan, and Shenzhen Stock Connect sold 120 million yuan and bought 249 million yuan.

On the news, on the evening of August 14, Liu Qingfeng, chairman of iFLYTEK, sold 39.9561 million shares of his company, and his shareholding ratio was reduced from 7.27% to 5.54%. Although iFLYTEK explained in the announcement that the reduction was to repay the debt left over from the previous increase, it still affected market sentiment to some extent. In addition to the chairman, China Mobile, the largest shareholder, also announced at the end of July that it would reduce its stake in iFLYTEK by 1% between August 22 and December 31.

At the same time, the overall revenue data of iFLYTEK's half-year report released on August 14 was not impressive, and the net profit fell in a state of double decline, with the net profit attributable to the parent falling by 73.54% year-on-year, and the asset-liability ratio also reaching 49.83%, the highest in history.

Not only that, the delivery difficulties at the end of 2022 also affected the revenue growth of iFLYTEK's education business, which increased by 2.56% year-on-year in 2022 for education products and services, but reported a year-on-year growth rate of 27.38% in the first half of 2022. Although the automobile, financial, medical and other businesses have grown, the proportions of these businesses to the total revenue in the first half of 2023 are 2.72%, 1.16% and 2.45%, respectively, which is a limited proportion.

For the current iFLYTEK, the new growth points brought by the big model are indispensable. In Liu Qingfeng's words, for companies without technology accumulation, it is a reasonable choice to do vertical large models to solve industry application needs, but for iFLYTEK, doing a good job of general large models may have a place in the future.

From the perspective of the capabilities of the general large model itself, after the goal of fully benchmarking ChatGPT on October 24, iFLYTEK still needs to face the inquiry of "when can it benchmark GPT 4.0". In this regard, iFLYTEK is facing the bottleneck of computing power and the problem of ecological construction.

The biggest obstacle to increasing the amount of parameters is still computing power. Liu Qingfeng mentioned in the group interview that the original computing power of about 300 to 400P can make a 100-billion-level model very well and build industry applications, but the parameter amount of GPT4.0 reaches 1.3 trillion, which requires at least a few thousand P computing power. According to him, the construction of this part of the computing power will be completed in October this year.

As for the scale of users, due to policy restrictions, domestic large-model products have not yet been officially opened for commercial use. With the official implementation of the Interim Measures for the Management of Generative Artificial Intelligence Services, it is unclear whether this situation will change. This is also one of the purposes of iFLYTEK's ecosystem - to expand the user base as much as possible and obtain training data.

How long it takes for the commercial value of large models to be reflected in earnings reports is obviously a topic of greater concern for investors. In this regard, the recent semi-annual report released by iFLYTEK has not performed well. After being equipped with the capabilities of large models, the revenue growth rate of C-end education products represented by iFLYTEK AI learning machines is obvious, but in the more closely connected industries in the automotive, financial, medical and other businesses, the large model has not shown a pulling effect.

From 2019 to 2022, iFLYTEK's government subsidies included in the current profit and loss were 412 million, 426 million and 439 million, accounting for 50.24%, 31.23% and 28.21% of the total profit, respectively. This downward trend was reversed in 2022, and the financial report showed that the government subsidy included in the current profit and loss in 2022 was 1 billion yuan, but the net profit attributable to the parent company was 561 million yuan.

As a strategically important pedestal business, the business value brought by the large model is often difficult to measure directly. However, for iFLYTEK, whose revenue and profit fell in the first half of the year and its profitability was questioned, the ability improvement and commercial implementation of the Xinghuo model are particularly important. At the press conference of version 2.0, iFLYTEK further demonstrated its ambition to build an ecosystem, but it remains to be seen whether it can be effective in the future.