On January 30, iFLYTEK held a press conference for the upgrade of the Xinghuo cognitive model V3.5, and officially launched the iFLYTEK Xinghuo V3.5 based on the first national computing power training.
With the sweep of the wave of large models, major manufacturers have begun to expand their layout in the field of large models, and iFLYTEK has also responded positively. On October 24, 2023, iFLYTEK and Huawei jointly announced the official launch of the first Vanka domestic computing platform "Feixing-1" to support the training of trillion-parameter large models. In the more than 90 days after its launch, iFLYTEK Xinghuo continued to increase R&D investment, and based on the "Feixing No. 1", it launched a large-scale model training with larger parameters against GPT-4, laying the foundation for the upgrade and release of iFLYTEK Xinghuo V3.5 on January 30.
Some of the capabilities have caught up with GPT-4
At the latest iFLYTEK press conference, the company highlighted a series of eye-catching keywords, including "surpassing GPT-4" and "domestically produced independent and controllable computing platform". The Xinghuo large model V3.5 released this time has been fully upgraded, and its performance is not only close to the level of GPT-4 Turbo, but also has made significant breakthroughs in several key areas.
It is understood that after the Xinghuo large model V3.5 has been fully upgraded, its performance has been close to the level of GPT-4 Turbo. Specifically, it has surpassed GPT-4 Turbo in language understanding and math capabilities, with 96% of GPT-4 Turbo's code capabilities and 91% of GPT-4V's multimodal comprehension capabilities.
Liu Qingfeng, Chairman of iFLYTEK, said: "The capability improvement of iFLYTEK Xinghuo V3.5 has reached a critical turning point. He predicts that by 2024, the iFLYTEK Xinghuo cognitive model will show excellent performance in more scenarios and fields.
This series of developments has raised people's attention to the core of the development of large models.
First of all, it is worth noting that the Xinghuo large model V3.5 has surpassed GPT-4 Turbo in terms of language understanding and mathematics. This indicates that the development of large models is moving towards a more comprehensive and in-depth direction. Language understanding has always been one of the core concerns in the development of large models, and iFLYTEK's new model has performed impressively in this regard. This means that in the future of natural language processing tasks, iFLYTEK's large model is expected to play a more important role, providing more accurate and intelligent solutions for various application scenarios.
Secondly, in terms of code capabilities, the Xinghuo large model V3.5 has reached 96% of GPT-4 Turbo. This reflects a significant improvement in the ability of large models to understand and generate code. As information technology continues to evolve, so does the need for large models with strong code understanding and generation capabilities. The superior performance of iFLYTEK's new model in this field indicates the application potential of the large model in promoting software development and automated programming.
Finally, the multimodal comprehension capability is another highlight of the iFLYTEK Xinghuo large model V3.5, reaching 91% of GPT-4V. This means that the model is better able to understand and process multiple input data, including text, images, sounds, and other modalities. This is of great significance for realizing more intelligent and integrated human-computer interaction, information processing and other applications.
Overall, the Spark model V3.5 presented by iFLYTEK in this press conference not only surpasses GPT-4 Turbo in terms of performance, but also makes significant progress in several key aspects. These breakthroughs mark the evolution of the core of the development of large models, providing more powerful and intelligent solutions for all walks of life. At the same time, as a large model of domestic independent and controllable, iFLYTEK's achievements also highlight China's increasingly strong position in the field of artificial intelligence.
However, it is worth noting that more independent assessment and verification of the objectivity and reliability of these claims is still needed. In the highly competitive tech sector, objective data and evaluations will help better understand the real value of this new model.
Promote the development of large models
According to reports, the seven core capabilities of Xinghuo V3.5 have been comprehensively improved, including text generation increased by 7.3%, language comprehension increased by 7.6%, knowledge question answering increased by 4.7%, logical reasoning increased by 9.5%, mathematical ability increased by 9.8%, code ability increased by 8.0%, and multimodal ability increased by 6.6%.
The Xinghuo V3.5 launched by iFLYTEK has achieved comprehensive improvements in text generation, language understanding, knowledge question answering, logical reasoning, mathematical ability, code ability, and multimodal ability.
First of all, the improvement of text generation is one of the core capabilities of large models. In the age of information explosion, it will be an important task for models to generate more accurate and expressive text. Xinghuo V3.5 achieves a 7.3% improvement in text generation, which provides strong support for the model to better understand and generate natural language. This is closely related to the development trend of large models in natural language processing tasks.
Secondly, the improvement of language comprehension is also an important direction for the development of large models. Xinghuo V3.5 has achieved a 7.6% improvement in language understanding, showing that the ability of large models to understand context and reason about semantic relationships has been continuously enhanced. This has a positive effect on the realization of tasks such as more intelligent dialogue systems, sentiment analysis, etc.
Knowledge question answering and logical reasoning are the other two key directions in the development of large models. Xinghuo V3.5 achieves 4.7% and 9.5% improvement in these two aspects, respectively, indicating that the performance of large models in dealing with complex problems and logical reasoning is constantly improving. This is of great significance for solving complex problems in the real world, such as intelligent customer service, legal consultation, etc.
The improvement of mathematical ability and code ability provides a broader space for the application of large models in the field of science and technology. Xinghuo V3.5 has achieved 9.8% and 8.0% improvement in these two aspects, respectively, providing more reliable support for the model to better handle mathematical problems and generate code. This is of great significance for promoting the application of large models in the field of engineering.
Finally, the improvement of multimodal capability provides a better solution for large models to process multiple information sources such as images and voices. Xinghuo V3.5 achieves a 6.6% improvement in multimodal capabilities, which provides strong support for the model to better understand and process multimodal information. This is an important impetus for the realization of more comprehensive and complex human-computer interaction systems.
Overall, the seven core capabilities of Xinghuo V3.5 demonstrate the wide application potential of large models in different fields. The future development trend of large models will mainly focus on text generation, language understanding, knowledge question answering, logical reasoning, mathematical ability, code ability, and multimodal ability. However, the field of large models still needs to pay attention to issues such as transparency, fairness, and data privacy, so as to balance technological innovation and ethical responsibility, and promote AI technology to better serve society.
A large model of national openness
With the rapid development of science and technology, open large models are becoming one of the important engines to promote innovation in the field of artificial intelligence.
The open model makes AI technology more popular and democratized. Ordinary users can use AI technology more conveniently and enjoy more intelligent services and experiences through stronger voice interaction, text understanding, and multimodal capabilities. At the same time, the use of open large models promotes innovation in various industries, especially in the fields of customer service, education, healthcare, and entertainment. Stronger model capabilities mean more efficient and personalized services, which further promotes the digital and intelligent development of the industry.
First of all, in terms of industry applications, the launch of the national open model will have a far-reaching impact on all walks of life. In the fields of customer service, automobiles, and robots, human-computer interaction will be more intelligent and natural. The upgrade of language understanding, text generation, and knowledge Q&A of the open large model will help the industry achieve more efficient and intelligent services and communication, and promote intelligent transformation.
Secondly, in the field of education, the application of open large models in the field of education will provide students with a more personalized and efficient learning experience. Through voice interaction, knowledge questions and answers, students can obtain knowledge more conveniently and improve learning efficiency. Educational institutions and platforms can use open models to provide customized teaching content to facilitate the popularization and dissemination of knowledge.
Then, in the research field, researchers will benefit from the improvement of open large models in text generation, logical reasoning, etc. This will help accelerate scientific discovery and innovation, making it easier for researchers to access and process large volumes of literature and information. The Open Model provides a more powerful tool for interdisciplinary research and promotes deeper progress in various fields of the scientific community.
Finally, in terms of social communication, in terms of multilingual support, the upgrade of the open model will promote cross-cultural communication and understanding. The upgrade of the iFLYTEK translator will provide users with a more free and natural language communication experience, which is expected to reduce language and cultural differences and promote the process of a globally interconnected society.
Overall, the launch of the Open for All model marks the popularization and application of AI technology around the world. The wide application in different fields will promote industrial upgrading, promote educational innovation, accelerate scientific research, promote social exchanges, and bring more possibilities to human society. By continuously improving the performance of the open model, iFLYTEK is making an important contribution to building a smarter and more interconnected future society.
However, its development has not only brought many positive meanings, but also faced a series of challenges.
Data privacy and security: As the use of models becomes more widespread, more attention needs to be paid to the privacy and security of user data. It is an important challenge to ensure that users' personal information is adequately protected during the use of the open model.
Computing power and energy requirements: Training large-scale models requires a huge amount of computing power and energy investment, which may have a negative impact on the environment. Developers and researchers need to focus on the sustainability and environmental friendliness of model training while pursuing performance.
Transparency and explanatory: As models become more complex, their decision-making processes become more difficult to understand. For open large models, improving their transparency and explainability is key to ensuring user trust and control.
Legal and ethical issues: In the use of the large model of openness for all, legal and ethical issues may be involved, such as intellectual property rights, allocation of responsibilities, etc. Relevant regulations and ethical standards need to be further improved to ensure the legal and compliant use of the model.
The Open for All model has provided a huge boost to the advancement of AI technology, but in the process of solving these challenges, all parties need to work together to ensure the healthy development of this technology and bring more benefits to society.
What is the overall strength of the iFLYTEK model?
In the field of artificial intelligence, iFLYTEK has always been one of the leading companies that has attracted much attention. The company has made remarkable achievements in speech recognition, natural language processing, and other fields. Among them, its large model technology has always been an important part of the trendset. However, to fully assess the overall strength of the iFLYTEK model, it is necessary to conduct a comprehensive analysis of its performance in multiple aspects.
In terms of scientific research strength, iFLYTEK has strong strength in artificial intelligence research, and continues to promote cutting-edge research in the field by publishing papers at top international conferences. According to the data released by it in the third quarter of 2023, iFLYTEK recorded a revenue of 12.614 billion yuan in the first three quarters, and the revenue scale decreased slightly by 0.37% compared with 12.661 billion yuan in the same period last year. The net profit attributable to shareholders of listed companies also declined sharply, and the net profit attributable to shareholders of listed companies of 99.36 million yuan decreased by 76.36% year-on-year.
iFLYTEK explained in the financial report that the main reason is that the company actively seizes the new historical opportunities of general artificial intelligence and firmly invests in general artificial intelligence cognitive models.
In terms of technical strength, iFLYTEK's large model has a deep background in neural network and machine learning, and uses a large amount of data for training to continuously improve the generalization ability and adaptability of the model. This makes it excellent for dealing with a wide range of language variants and accents. However, the basic ability is slightly weaker, in August last year, a large model experience report released by the Xinhua News Agency Research Institute showed that Baidu Wenxin Yiyan is the leading level in China in terms of basic capabilities of large models, and the advantages of the Spark large model are manifested in work efficiency and commercial applications.
In terms of application, although the Xinghuo model is a general model, iFLYTEK also anchored many application scenarios for it at the press conference, but it did not fall into the commercialization circle of the general model, but implanted it into the consumer products represented by AI learning machines at the first time.
As more and more players run into the market, large models may not be able to support high premiums, and the profits of intelligent education hardware are bound to return to a reasonable range, and may even roll out the Internet genre that does not sell hardware but only software. At that time, iFLYTEK, which has shallow Internet genes, may suffer a lot of impact on its education fundamentals.
On the other hand, there is no so-called "technical myth" in the large model track, and many scenarios and applications need the support of underlying computing power. Although iFLYTEK is not afraid of players such as the future and homework gang in the short term, from a long-term perspective, if giants such as Baidu, Alibaba, and Tencent go deep into the battle, it may be difficult for iFLYTEK to have the ability to confront them head-on.
However, iFLYTEK has been improving its capabilities in all aspects of the large model, and in the future, with the continuous development of technology, it is believed that the large model of iFLYTEK will be further improved in continuous iteration to provide better support for a wider range of application scenarios.