Domestic large models are facing a peak moment.

From the perspective of conflicts, OpenAI's disconnection will directly stir up changes in the seating arrangements of the large model landscape; from the perspective of the industry's overall environment, the acceleration from general large models to vertical implementations has also led to intensified competition in niche positions.

Observing the development of domestic large models along these two directions, iFLYTEK Spark has stood out.

Especially, the release of iFLYTEK Spark V4.0, with an overall performance surpassing GPT-4 Turbo, indicates that the landscape of domestic large models is undergoing changes.

1. Release of the most powerful domestic model

iFLYTEK Spark V4.0 is here!

On June 27, iFLYTEK released the iFLYTEK Spark V4.0 large model and related implementations in Beijing. The seven core capabilities of iFLYTEK Spark V4.0 have been comprehensively improved, surpassing GPT-4 Turbo overall, ranking first in 8 international mainstream test sets, and leading domestic large models.

Based on the country's first domestic 10,000-card computing power cluster "Feixing No. 1", iFLYTEK Spark V4.0 was officially released. iFLYTEK Spark V4.0 has been comprehensively upgraded in seven core capabilities, fully benchmarked against GPT-4 Turbo, and achieved overall surpasses in text generation, language understanding, knowledge question answering, logical reasoning, mathematical abilities, and other aspects.

iFLYTEK Spark V4.0 has further upgraded its image and text recognition capabilities, and its application effects in scenarios such as scientific research, finance, healthcare, justice, and office work have surpassed GPT-4o.

It is worth noting that regarding the market's more concerned long-text processing capabilities, Spark has also been upgraded, and for the illusion problem of long-document knowledge question answering, it has pioneered a traceability function in the industry.

Interestingly, this year's middle school entrance and high school exams have become a touchstone for testing large models. In the actual test of this year's middle school entrance and high school exams, iFLYTEK Spark's "scores" in various subjects such as Chinese, mathematics, and foreign languages all ranked first, earning it the title of "the large model that is better at solving problems".

External authoritative test sets also reflect the leadership of iFLYTEK Spark V4.0.

Among 12 mainstream test sets for large models at home and abroad, iFLYTEK Spark ranked first in 8 test sets, surpassing international large models such as GPT-4 Turbo, and leading domestic large models overall.

On-site, Liu Qingfeng, Chairman of iFLYTEK, demonstrated the effects of iFLYTEK Spark V4.0 in complex instructions, complex logical reasoning, spatial reasoning, high school mathematics, and other aspects, with Spark's "intelligence" evolving once again.

Taking spatial reasoning as an example, in the on-site demonstration question, "Bob is in the living room. He takes a cup and walks to the kitchen. He puts the ball into the cup and then takes the cup to the bedroom. He turns the cup upside down and walks to the garden. He puts the cup in the garden and then walks to the garage. Question: Where is the ball?" iFLYTEK Spark can infer that the ball is on the floor of the bedroom based on space and common sense. The progress of these abilities is meaningful for future embodied intelligence and home robots.

2. Accelerated vertical implementations

As mentioned above, domestic large models are now accelerating their development in two directions, one being the overall performance of general large models and the other being the implementation capabilities of large models.

In terms of application scenarios, iFLYTEK Spark V4.0 has also continuously brought surprises to the market.

At the press conference, iFLYTEK Spark released and upgraded multiple products.

iFLYTEK Spark APP/Desk has been upgraded, releasing "Personal Space" to create an AI assistant that understands you better; for personalized applications in professional fields, iFLYTEK has upgraded the iFLYTEK XiaoYi APP, launched personal digital health space, and created a health assistant for everyone; the industry's first Spark intelligent grading machine, "AI Teaching Assistant" helps teachers reduce workload and increase efficiency, teaching students according to their aptitudes; iFLYTEK AI Learning Machine has upgraded its 1-on-1 Q&A tutoring function to create an AI learning assistant for every child.

iFLYTEK Spark APP took the lead in releasing "Personal Space", allowing users to upload their work, study, life, health, and other types of materials to form everyone's exclusive knowledge base, and then combine with personal settings to allow the large model to generate more personalized content. In addition, iFLYTEK Spark has launched the first batch of 14 intelligent agents, creating exclusive assistants for specific scenarios.

The capabilities of the iFLYTEK XiaoYi APP continue to be upgraded, covering 1600 common diseases, 2800 common drugs, and 6000 common examinations and tests, meeting users' core scenario health needs before seeing a doctor, during medication, and after examinations. Currently, the iFLYTEK XiaoYi APP has a cumulative download volume of 12 million, with a user satisfaction rate of 98.8% and an active recommendation rate of 42%.

The Spark intelligent grading machine integrates intelligent grading, precise learning conditions, and personalized learning. It supports free typesetting and homework of any paper size. While supporting intelligent grading of multiple subjects and question types, it can also generate multi-dimensional learning condition reports in real-time, providing materials for teachers' homework commentary and face-to-face tutoring.

On-site, the entire process of grading homework using the Spark intelligent grading machine was demonstrated, with 15 student assignments graded in half a minute. The grading simulates real handwriting, almost the same as teachers grading homework normally.

In addition, the Spark large model has also connected the entire ecosystem of iFLYTEK's C-end software and hardware products, allowing millions of smart hardware users to have the "Spark full set" with one click.

For example, files from iFLYTEK's smart office notebook and smart voice recorder can be synchronized to the Spark personal space with one click. Through data interoperability and operational linkage, synchronizing meeting records from an office notebook to Spark allows Spark to write official documents, make PPTs, and generate to-do items, bringing a more efficient office experience.

3. Accelerated commercialization

As everyone knows, iFLYTEK started with AI voice processing and, after years of development, is at the international top level in intelligent voice processing.

Recently, iFLYTEK's "Key Technologies and Industrialization of Multilingual Intelligent Voice" project, as the first completion unit, won the First Prize of the National Science and Technology Progress Award. At the press conference, the national award winner once again made a "big splash", and the Spark voice large model ushered in a new breakthrough.

Of course, iFLYTEK not only has long-term technological accumulation but has also been mature in commercial applications for a long time.

The first is the field of human-computer interaction.

With the support of the large model, the Spark intelligent cockpit has been upgraded, not only with "free interaction" in multiple languages and dialects but also with ultra-human-like interaction with multiple emotions and modalities, making human-vehicle interaction more warm. Currently, iFLYTEK's voice interaction products have a stable domestic market share of first place and are widely exported to various parts of the world. The Spark large model has given highly intelligent interactive experiences to many models from automakers such as FAW, Chery, GAC, JAC, and Great Wall.

To better implement the large model, iFLYTEK has also created solutions integrating cloud, edge, and terminals, as well as software and hardware, empowering more industry scenarios such as home appliances, operators, and robots. Responding to the needs of embodied intelligence and humanoid robot companies, iFLYTEK officially released the Robot Superbrain Platform 2.0 this time, the first in the industry to support multimodal interaction. Currently, over 400 robot companies have adopted the iFLYTEK Robot Superbrain Platform.

The second is that since the release of iFLYTEK Spark, the efficiency it brings to enterprise production has been recognized by the market.

Since its release on May 6 last year, iFLYTEK Spark has become the preferred choice for many leading enterprises in various fields such as the National Energy Group, China National Petroleum Corporation, China Mobile, PICC, China Pacific Insurance, Bank of Communications, Chery Automobile, FAW, Volkswagen, JAC Group, Haier Group, and Midea Group.

iFLYTEK Spark has already produced application results in multiple typical scenarios such as code, compliance review, customer service, bid evaluation, and intelligent interaction. Taking Bank of Communications as an example, the product iFlyCode based on Spark's large model capabilities covers over 6,000 R&D personnel, with a code adoption rate of 38%, significantly improving work efficiency.

At this press conference, the Spark Enterprise Intelligent Agent Platform was officially released. Focusing on the three key capabilities of building intelligent agents, the current Enterprise Intelligent Agent Platform has covered over 400 AI atomic capabilities, integrated over 90 external information sources, and connected over 100 internal IT systems, allowing enterprises to quickly build implementable intelligent agent applications in combination with business scenarios. The platform has also launched 32 enterprise intelligent agents in the production, scientific innovation, office, and management domains, which are ready for enterprises to plug and play.

For example, in the code intelligent agent iFlyCode, it integrates six scenario intelligent agents such as code generation assistant, architecture design assistant, code question answering assistant, testing assistant, database optimization assistant, and code review assistant, increasing the adoption rate from 30% to 52%, significantly enhancing the practicality of enterprise intelligent agents.

While bringing efficiency improvements to the industry, iFLYTEK Spark's large model also contributes to the large model ecosystem.

Since the release of iFLYTEK Spark V3.5 on January 30 this year, in just five months, the Spark developer ecosystem has accelerated its growth, with the number of developers increasing from 5.98 million to 7.02 million, an increase of over 1.04 million; the number of overseas developers exceeds 400,000; and the number of large model developers reaches 570,000. More and more developers are joining the Spark ecosystem, unleashing more application value in rigid demand scenarios.

Liu Qingfeng said that only an independently controllable and prosperous ecosystem can lead to a great future for China's general artificial intelligence.

Facing the new ecosystem of artificial intelligence in the future, he emphasized the need to focus on the source technology ecosystem, intelligent agent ecosystem, application ecosystem, and industry ecosystem, achieving independent controllability and integration of software and hardware to achieve deep implementation of large models; it is necessary to scientifically and rationally understand the comprehensive gap between China and the United States in large models, and also have confidence to catch up quickly, providing a complete set of strategies from source technology to industrial ecology to application implementation, building a truly independently controllable AI industrial ecosystem with a long-term perspective.