Domestic large models reach new heights! iFlytek Spark 4.0 released: overall surpassing GPT-4 Turbo, ranking first in 8 international authoritative test sets

The capabilities of domestic large models have reached a new level!

On June 27, iFlytek officially released the iFlytek Spark model V4.0, as well as artificial intelligence applications in many fields such as medical care, education, and business.

With the release of the new version, the seven core capabilities of iFlytek Spark V4.0 have been fully upgraded, ranking first in 8 international mainstream test sets, overall surpassing GPT-4 Turbo, and leading domestic large models.

Liu Qingfeng said that currently, the number of downloads of Spark APP has reached 131 million, and a number of user-favorite application assistants have emerged. With the support of the Spark model, the sales of smart hardware in some scenarios increased by 70%+ year-on-year, and the average monthly usage exceeded 40 million.

In addition, the Spark V4.0 large model is trained based on the country's first domestic Wanka computing cluster "Flying Star One", which means it is completely autonomous and controllable.

Overall surpassed GPT-4 TurboSeven underlying core capabilities have been upgraded again

At the end of January this year, Spark Large Model V3.5 surpassed GPT-4 Turbo in terms of language understanding and mathematical capabilities, but the code , multi-modality and other capabilities still have some gaps with the latter.

Now, 5 months have passed, Spark Model V4.0 has evolved again, not only completing the overall surpassing of GPT-4 Turbo in the five major capabilities of text generation, language understanding, knowledge question and answer, logical reasoning and mathematics , and further narrowed the gap in code and multi-modal capabilities, especially the multi-modal capability reaching 97% of the latter level.

Not only that, in 12 mainstream Chinese and English test sets covering comprehension & reasoning, comprehensive examinations, mathematics & science, and coding tasks at home and abroad, Spark Model V4.0 ranked first in 8 test sets.

In addition to the mainstream test sets, the recently released high school and college entrance examination questions can better test the comprehensive quality of large language models.

In the assessment in this field, the performance of Spark Large Model V4.0 was particularly outstanding. Taking the 2024 Beijing High School Entrance Examination as an example, Spark Model V4.0 achieved first place in both subjective and objective question score rates, which can be said to be a qualified or even average-level candidate.

In addition to the comprehensive improvement of the seven core capabilities of the base, the Spark Model V4.0 has also carried out targeted performance optimization and functional innovation in many aspects such as command following, text, multi-modality, and reasoning capabilities. .

First of all, Spark Large Model V4.0 strengthens complex command following and long text processing capabilities, and is the first in the industry to feature a new feature - long text content traceability.

Specifically, in tasks such as long document feature extraction, long document summary summarization, long document Q&A, and long document text generation, the overall performance of Spark Large Model V4.0 is comparable to GPT-4 Turbo. At the same time, the content tracing function further reduces illusions in long document knowledge question and answer tasks, making the answers more accurate.

In terms of multi-modality, Spark’s image and text recognition capabilities continue to upgrade, especially in professional fields such as scientific research, finance, medical care, justice, and office. The image and text recognition capabilities have been greatly improved, surpassing the capabilities of OpenAI released last month. The latest flagship model GPT-4o. This means that Spark Large Model V4.0 will have greater application potential in these vertical application fields in the future.

Starfire Large Model V4.0 is also more capable in image and text recognition tasks for complex educational scenes, and significantly surpasses GPT-4o in complex formula recognition of printed and handwritten characters.

Similarly, in terms of multi-modal understanding based on logical relationships, Spark Model V4.0 can give answers with more rigorous logic and clearer ideas than before.

Finally, Spark Model V4.0 can solve more complex logical reasoning and spatial reasoning problems.

Take spatial reasoning as an example, “Bob was in the living room. He walked to the kitchen with a cup. He put the ball in the cup and then walked to the bedroom with the cup. He turned the cup upside down and then walked to the garden. He placed the cup in the garden and walked to the garage. Question: Where is the ball? "iFlytek Spark can infer that the ball is on the floor in the bedroom based on space and common sense. The advancement of these capabilities will be of great benefit to future embodied intelligence and homes. Robots all have meaning.

It can be said that the upgrade of a series of underlying core capabilities shows that domestic large language models comprehensively surpass foreign competing products, and lays a solid foundation for the implementation of applications based on large models.

Xinghuo Voice Model releases "Free Dialogue" in 74 languages and dialectsSolving speech recognition problems in strong interference scenarios

Voice capability has always been iFlytek’s absolute advantage. On June 24, iFlytek won the first prize of the 2023 National Science and Technology Progress Award for its "Multi-lingual Intelligent Voice Key Technology and Industrialization" project.

As early as January 30, 2024, iFlytek has released the Spark voice model for the first time in the Spark model V3.5 update. The speech recognition effect of the first batch of 37 mainstream languages exceeds that of OpenAI Whisper V3. At the launch of Spark V4.0, iFlytek announced that its speech model capabilities have been upgraded again. In addition to 37 mainstream languages, it also added recognition of 37 dialects. Users can realize free communication in 37 languages + 37 dialects, a total of 74 languages without cutting.

At the scene, the demonstrator directly communicated with the large model in Shanghainese, Cantonese, Hefei dialect, Sichuan dialect, Japanese, French and other languages, and the large model could be quickly and accurately recognized.

^{Live dialect recognition demonstration}

In addition, iFlytek also highlighted its ability to transcribe speech in ultra-complex scenes. The three people conducted an on-site measurement in a noisy scene where the simultaneous aliasing of speech was difficult for normal human ears to hear clearly. The multi-modal capabilities of iFlytek Spark not only enabled the role separation of three people's overlapping voices, but also real-time conversion. Write what each person says.

With the upgrade of intelligent voice, iFlytek has further deepened its presence in the automotive scene. Liu Qingfeng also showed off the upgraded Xinghuo smart cockpit.

Medical large model "iFlytek Xiaoyi"Everyone's AI health assistant

Due to the particularity of medical scenarios, extremely high professionalism and extremely low error tolerance, C-side health knowledge Q&A has always been is a difficult point.

In October last year, iFlytek briefly introduced its large-scale medical model application "iFlytek Xiaoyi" and its B-side application "Smart Medical Assistant" during the release of Spark V3. The C-side is open and can provide functions such as physical examination report analysis, and the "intelligent medical assistant" can provide pre-inquiry and other capabilities.

This time Spark V4.0 is released, iFlytek has further upgraded and introduced the large medical models and applications. The data shared by Liu Qingfeng at the scene showed that iFlytek Spark Medical’s core capabilities surpassed GPT-4 Turbo and GPT- 4o.

On the application side, since its release, the "intelligent medical assistant" for doctors has achieved 820 million auxiliary diagnoses, 1.47 million corrected diagnoses, and helped discover 72.67 million unreasonable prescriptions.

The personal-oriented App "iFlytek Xiaoyi" can provide ordinary users with free interpretation of medical records, physical examination reports, and examination reports, and can also conduct conversations about medical knowledge. By gathering all kinds of health information, the App can provide users with a personal digital health space that records disease history, medication history, living habits, etc. It can help users analyze the causes of illnesses before seeing a doctor, and provide users with personality judgment and drug contraindications when taking medication. , provide analysis and records of changes after inspection.

Live demonstration: Scan the QR code to upload medical records and physical examination reports, analyze and interpret them.

Release of intelligent marking machineThe large education model has been upgraded again

Iflytek Spark V4.0 has upgraded the large education model, and released two models of Spark intelligent marking machine and further upgraded iFlytek AI learning machine hardware.

The intelligent marking machine is aimed at teachers. Put the test paper on the marking machine. The marking machine can automatically scan and print and mark the original paper, and the scanned data will be automatically uploaded to generate an analysis of the student's personality and class status. Commonality analysis can also be used to assign individual assignments to students. Compared with manual marking, the efficiency of intelligent marking machines in marking scenarios has been increased from 90 minutes per class to 5 minutes per class.

In addition, iFlytek has also continued to upgrade its educational hardware "AI learning machine". Using the upgraded "AI learning machine", users only need to take pictures of test papers and select test questions, and AI will help users answer questions and provide intelligent conversational explanations.

^{Launching "Personal Space"Creating an AI assistant that understands you}

The popularity of "intelligent agents" has become an important sign of the explosion of applications in 2024.

Since OpenAI launched GPTs and GPT Store, people can tailor-make their own AI smart assistants based on large models. Many people have great expectations for their arrival, believing that it will usher in the "iPhone moment" of AI. Other manufacturers have followed suit and launched similar AI agent services. For example, Microsoft has launched its own Copilot GPTs service.

However, things did not develop as expected, and GPTs quickly ran into a bottleneck. A few days ago, Microsoft announced that it was cutting off the Copilot GPTs service, which had lasted only three months. A closer look at the reasons shows that GPTs has largely failed to keep up in terms of application scenarios and commercial implementation, and has subsequently lost momentum for further development.

Therefore, in the past six months, domestic large-scale head model manufacturers have been rapidly developing intelligent bodies and have made great efforts at the application level.

This iFlytek Spark V4.0 version also focuses on the launch of "intelligent" updates. In iFlytek Spark Desk and Spark App, "Agent" has become the same first-level entrance as Chatbot. Click on the "Agent Center" and iFlytek's smartphone store covers smartphones in multiple fields such as productivity tools, learning, programming, and marketing.

According to the information disclosed at the press conference, currently, Spark APP/Desk will launch the first batch of 14 intelligent agents to create exclusive assistants for specific scenarios.

Users can either choose their own AI English teacher here, or choose a customized AI lawyer to help them draft a contract. They can also choose AI programmers who are good at Python, C++ and other languages to help them. Implement programming yourself.

For example, choose a strict "IELTS teacher". After clicking on the agent creation, we can summarize the "IELTS teacher" we need in one sentence - he needs to help me develop a three-month study plan, covering listening, speaking, reading and writing, and teaching for each part.

After clicking Generate, the background will automatically expand other aspects of the agent according to needs, and you can debug it at any time.

After debugging is completed, you can learn from the "AI IELTS teacher".

However, compared with GPT Store, iFlytek regards the "intelligent" function as part of achieving user personalization. Users can choose and define the smart assistant they need according to their own needs.

In terms of realizing "personalization", iFlytek Spark V4.0 also released a "personal space" to provide users with an exclusive private domain knowledge base. By uploading personal documents, large models can be processed more accurately. Questions and answers and content generation; and through personal tagging, schedule management, information subscription, and speaker creation, it provides users with more personalized and interesting services.

Users can continue to upload their own information files in their personal space. AI will conduct questions and answers based on the uploaded information and provide content traceability, reducing model illusions while improving personalization.

By allowing users to define "Chatbot" to a higher level, from personal knowledge enhancement, to selecting personal tags, creating speakers, to open agent definition, iFlytek will use the data of the large model "Chatbot" Definition autonomy has been achieved in all aspects of , workflow and expression forms.

Personalized large models have reached a new stage. Spark Large Models V4.0 takes the matter of "everyone's personalized intelligent assistant" a step forward in both "personalization" and "intelligence".

Written at the end

In 2024, the first year of the big model’s launch, if you want to maintain a leading position, it is crucial to be the first and have all-round layout.

A few days ago, iFlytek's "Multi-lingual Intelligent Voice Key Technologies and Industrialization" project won the first prize of the National Science and Technology Progress Award. This may be one of the affirmations of iFlytek’s AI technology achievements over the years. You know, this is the first first prize in the National Science and Technology Progress Award in the field of artificial intelligence in the past ten years since the deep learning revolution.

It can be said that from V3.5 at the end of January to V4.0 today, the iFlytek Spark model has not only become more stable in terms of underlying capabilities, but has also been fully developed in terms of application implementation.

On the one hand, we are constantly improving the underlying core language capabilities, continuing to catch up with the world's top large models, and polishing and upgrading our own advantageous model capabilities such as voice capabilities; on the other hand, we are investing more energy in application and commercialization, and starting from reality. Based on the needs of the scene, a comprehensive layout is carried out, and various methods such as building vertical large models, privately customized agents, and end-side intelligent hardware are used to accelerate the implementation of large models on the B-side and C-side, so that more enterprises and ordinary users can experience the big data first-hand. The value the model brings.

The above is the detailed content of Domestic large models reach new heights! iFlytek Spark 4.0 released: overall surpassing GPT-4 Turbo, ranking first in 8 international authoritative test sets. For more information, please follow other related articles on the PHP Chinese website!