GuShiio.com 
Image source: Generated by AI
At the beginning of 2025, DeepSeek caused an earthquake in the domestic and foreign large model industries. In addition to the excellent performance of the deep reasoning model DeepSeek-R1 in answering questions, the existence of DeepSeek has injected a tense and lively air into the domestic model circle.
First, relying on its technological advantages, DeepSeek entered the head echelon of international large models in one fell swoop, which gave domestic large model companies the possibility of overtaking in corners.
Second, DeepSeek’s training results break the problem of limited computing power and prove that through algorithm optimization, high-quality models can also be trained with low computing power.
When the worry of “stuck computing power” is put down, what problems should a large model that takes into account energy consumption and accuracy solve? At this level, large domestic model companies have handed in their own answers.
Recently, Zhongke Wenge, an AI company incubated by the Automation Institute of China Academy of Sciences, released the flagship version of the Yayi Model–YAYI-Ultra gave its own answer before solving the dilemma of landing precision-energy consumption for large models.
As an authoritative evaluation system covering 100+ models around the world, the OpenCompass List has always been a barometer for observing the technical route of large models. In its recently released Open Academic List of OpenCompass Large Models, Zhongke Wenge YAYI-Ultra broke into the top ten for the first time with a score of 64.5, becoming one of the five major China models in the TOP 10.
In OpenCompass’s latest open academic real-time list of large language models, YAYI-Ultra ranks tenth with a comprehensive score of 64.5, among which:
Code generation:LiveCodeBench ranks fifth, outperforming GPT-4o-20241120 version
Understanding complex instructions:IFEval ranked ninth
Knowledge reasoning ability:MMLU-Pro ranked ninth
In the C-Eval review, which focuses on Chinese understanding, YAYI-Ultra ranks second in the list of public access that allows independent verification, demonstrating its technical advantages in Chinese scenarios.
First-hand measurement: ultra-long text output
Accurate control of complex task planning
According to official information, YAYI-Ultra has outstanding performance in chart understanding, complex tasks, and long text understanding and generation. We immediately started using six dimensions.(In-depth understanding of multi-modal charts, complex image understanding, intelligent complex task planning (Function Call), data statistical analysis, and ultra-long text understanding and generation)Measure how YAYI-Ultra performs.
01 Visual understanding is upgraded again: understand language, understand charts more
Come on, let’s read a chart and try it.
YAYI-Ultra can accurately identify different colors and numbers in the histogram, fully understand the chart and give answers.
In addition to Chinese scenarios, in multi-language scenarios, YAYI-Ultra can also accurately understand and follow user instructions, providing accurate cross-language responses.
prompt：How did the distribution of agriculture-related employment change between 2012 and 2022? Did it increase or decrease, and by what percentage or amount? Answer in Chinese。
It can be seen that in terms of visual understanding, for technical difficulties such as cross-language multimodal alignment, multi-graph reasoning, and variable resolution, YAYI-Ultra has been comprehensively upgraded, enhancing the model’s ability to operate.Cross-language chart understanding, multi-chart question and answer, multi-modal command followingWith the ability to easily deal with complex chart scenarios such as stacked bar charts, scatter charts, and mixed charts, it is also outstanding in tasks such as chart redrawing and chart conversion.
02 Form Smart Explanation: Thousands of forms are not a problem
At work, statistics on complex reports are time-consuming and labor-intensive. We “fed” YAYI-Ultra a table containing three alternating types of industry general reports, industry in-depth reports, and company general reports. YAYI-Ultra accurately calculates the number of different types of reports.
prompt: What is the number of reports per type?
When it comes to irregular tables, YAYI-Ultra can still accurately parse and extract key data. The following table contains the total score structure and complex data expression. YAYI-Ultra can accurately understand the model types, methods and local indicator change values in the table, and complete comparative analysis.
prompt: Which base model has the most localized decline after using the IKE method?
In terms of statistical data understanding, it can be seen that YAYI-Ultra focuses on enhancing the questionnaire Q & A.Complex typography understanding, cross-language Q & Aand other abilities.
From financial reports and academic papers to complex tables with nested structures, YAYI-Ultra can accurately locate information and understand user intentions; at the same time, the model can also provide efficient and clear answers in cross-language table question and answer scenarios.
03 Function Call: Intelligent Planning for Complex Tasks
Continue to improve the difficulty, let’s make YAYI-UltraDraw a broken line chart of the number of gold, silver and bronze medals for the China team at last year’s Olympic Games (changes over time)。
First of all, it can be seen that YAYI-Ultra accurately understood the user’s intention, determined that “last year’s Olympics” referred to the Paris Olympics, and formulated a detailed mission plan; next, the model obtained data related to the gold, silver and bronze medals of the China team in the Paris Olympics through a search engine (including the types and winning times of 91 medals); then the medal data was sorted out, grouped, sorted in time, and code was generated. By calling the code interpreter, the line graph was completed.
The reason why YAYI-Ultra can complete this series of complex tasks disassembly and planning is inseparable from its increased tool calling capabilities. It mainly includes basic tools such as search engines, code interpreters, image analysis, and weather; news hot list tracking, and communication influence analysis and other special vertical domain tools.
The model significantly enhances the planning rationality under multiple tool serial invocation scenarios, and also improves the information collection capabilities under complex search scenarios.
04 Multimodal output: Illustrated pictures and texts, intuitive and concise
In the process of literature reading or information collection, we usually need to find and analyze specific information (such as numerical changes, experimental results, etc.) from multiple documents. Now we can find the desired content in one sentence, and YAYI-Ultra is based on text analysis. Based on the description, the corresponding picture content can be given simultaneously.
For example, ask questions:The percentage of different behaviors under different collaborative strategies
Based on the questions, YAYI-Ultra identifies multiple related artificial intelligence papers from the “artificial intelligence paper knowledge base” built by the user and answers them accordingly. The answer not only contains text, but also gives the original picture at the corresponding reference position, which greatly improves the reading experience and the reliability of the answer.
05 Long article on the whole stack: swallow thousands of words and write like a god
The most eye-catching thing is ultra-long text output. YAYI-Ultra supports up to 200,000 words of input and 100,000 words of ultra-long text output, forming a full-link closed-loop of long-text capabilities from “input understanding” to “content creation”.
YAYI-Ultra supports dual modes of networked intelligent creation and document-anchored creation, and decomposes long-text writing tasks into smaller and more controllable sub-tasks (first generating outlines, and then generating full-text based on the outlines), which effectively ensures the text structure and improves the quality of long-text generation.
● Connected intelligent creation: Collect information online to complete creation
prompt: Write a 30000-word historical analysis report on the development of Confucian culture in China
● Document-anchored creation: Defining knowledge boundaries and precise writing
prompt: Please write a long article based on the reference materials, with the theme of “Universal Artificial Intelligence Solutions: The Perfect Combination of Innovation and Efficiency”
06 Data analysis: accurate solution, visual interaction
Finally, we also conducted actual measurements on basic data analysis and visual charting, and YAYI-Ultra accurately completed the analysis, calculation and charting tasks.
Prompt: Calculate the per capita monthly income according to the table, then calculate the difference between the monthly income and the per capita monthly income, and draw a bar chart. The horizontal axis is the name, the vertical axis is the difference, and the title is “The gap between per capita income and the average.”
According to user requirements, YAYI-Ultra uses Python of Thought (POT) capabilities to generate Python code and execute it, accurately completing numerically-intensive tasks such as statistical inference, matrix operations, and numerical optimization.
From “flood irrigation” to “precise matching”
YAYI-Ultra is configured with flexible experts
Break through the bottleneck of implementing large models
At present, the implementation of the AI model is facing a critical juncture in the widening of the “capability-cost” scissors gap.
According to the latest IDC report, enterprises face the problem that model accuracy cannot fully meet business needs during the implementation of large AI models; at the same time, 92% of enterprises believe that the lack of computing power resources is the biggest challenge during the engineering implementation stage of large models.
Zhongke Wenge Yayi’s technical team revealed that YAYI-Ultra is a hybrid expert model featuring multi-domain capabilities. In order to improve the performance of professional tasks in different fields, it adopts a flexible expert configuration model to support mathematics, code, finance, public opinion, traditional Chinese medicine, security and other fields. The combination of experts in various fields can significantly alleviate the “seesaw” phenomenon that is common in dense models during the vertical domain migration.It can provide high-precision, low-energy intelligent solutions for the industry based on the needs of different fields.
For example, in the media field, Zhongke Wenge launched the Hongqi 3.0 Media Intelligent Platform, which is based on YAYI capabilities to help customers shorten content creation time by 30%-50%, and increase content release frequency by 20%-40%. After introducing automated review capabilities, the content error rate has dropped from 5% to about 0.5%, and has now been widely used in leading media such as Xinhua Agency, CCTV, and China Daily.
Zhongke Wenge Hongqi 3.0 Integrated Media Intelligent Platform
In the medical field, the YAYI-based Big Medical Golden Chamber Traditional Chinese Medicine Model can accurately diagnose more than 500 common diseases and provide patients with personalized treatment plans. The accuracy rate of dialectical reasoning after evaluation by clinical experts is as high as 90%. It performed well in the simulation test of the Chinese medicine doctor qualification examination. The performance was excellent, with an accuracy rate of over 94%, and the “Big Medical Golden Chamber” Chinese medicine health management APP was launched for C-end users.
China Academy of Chinese Medical Sciences Zhongke Wenge Da Medical Jinkui Traditional Chinese Medicine Health Management APP
In the field of finance and taxation, based on YAYI’s fiscal and taxation knowledge model, after special evaluation, the model answer accuracy rate is 90.1%, which is higher than other models of the same type. After accessing the large model, customers can provide 24-hour uninterrupted consultation services, allowing users to reduce queuing time by about 50%, and user satisfaction has increased by more than 30%.
Aerospace Information and Zhongke Wenge jointly developed a fiscal and taxation knowledge model
Currently, YAYI-Ultra (yayi.wenge.com) has opened up data analysis, knowledge base document analysis, and ultra-long text writing functions on its official website. Interested friends can also log in and try them out.
