When the concern of computational power being a bottleneck is set aside, what problems should large models that balance energy consumption and accuracy address?

Source: Light Cone Intelligence

Image source: Generated by Wujie AI

At the beginning of 2025, DeepSeek caused a seismic shift in the domestic and international large model industry. In addition to the excellent performance of the deep reasoning model DeepSeek-R1 in answering questions, the existence of DeepSeek has injected a tense yet vibrant atmosphere into the domestic large model circle.

Firstly, with its technological advantages, DeepSeek has entered the top tier of international large models, showing domestic large model companies the possibility of overtaking on a curve.

Secondly, the training results of DeepSeek have broken the limitations of computational power, proving that high-quality models can also be trained with low computational power through algorithm optimization.

When the concern of computational power being a bottleneck is set aside, what problems should large models that balance energy consumption and accuracy address? On this level, domestic large model companies have submitted their respective answers.

Recently, the AI company Zhongke Wenge, incubated by the Chinese Academy of Sciences, released the flagship version of the YAYI large model—YAYI-Ultra, which provides its own answer before breaking the "accuracy-energy consumption" dilemma of large model implementation.

As an authoritative evaluation system covering over 100 models globally, the OpenCompass ranking has always been a "barometer" for observing the technological routes of large models. In its recently released OpenCompass public academic ranking of large models, Zhongke Wenge's YAYI-Ultra scored 64.5 points, entering the top ten for the first time, becoming one of five Chinese large models in the TOP10.

In the latest OpenCompass real-time academic ranking of large language models, YAYI-Ultra ranked tenth with a comprehensive score of 64.5, including:

Code Generation: LiveCodeBench ranked fifth, outperforming the GPT-4o-20241120 version.

Complex Instruction Understanding: IFEval ranked ninth.

Knowledge Reasoning Ability: MMLU-Pro ranked ninth.

In the C-Eval assessment, which focuses on Chinese understanding, YAYI-Ultra ranked second in the publicly accessible list that allows for self-verification, showcasing its technical advantages in Chinese scenarios.

First-Hand Testing: Ultra Long Text Output

Precise Planning for Complex Tasks

According to official information, YAYI-Ultra excels in chart understanding, complex tasks, long text understanding, and generation. We immediately tested YAYI-Ultra's performance from six dimensions (multimodal chart deep understanding, complex image understanding, intelligent planning of complex tasks (Function Call), data statistical analysis, and ultra-long text understanding and generation).

01 Visual Understanding Upgrade: Understanding Language and Charts Better

Let's start by reading a chart.

Prompt: In the years around 2000, which price range of property fees saw the most significant change in proportion?

YAYI-Ultra can accurately identify different colors and numbers in the bar chart, fully understanding the chart and providing an answer.

In addition to Chinese scenarios, in multilingual contexts, YAYI-Ultra can also accurately understand and follow user instructions, providing precise responses across languages.

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.

As seen, in terms of visual understanding, YAYI-Ultra has undergone a comprehensive upgrade to address technical challenges such as cross-language multimodal alignment, multi-chart reasoning, and variable resolution, enhancing the model's capabilities in cross-language chart understanding, multi-chart Q&A, and multimodal instruction following. It can easily handle complex chart scenarios such as stacked bar charts, scatter plots, and mixed charts, and also performs well in tasks like chart redrawing and conversion.

02 Intelligent Table Understanding: Handling Thousands of Tables with Ease

In the workplace, complex report statistics are time-consuming and labor-intensive. We "fed" YAYI-Ultra a table containing alternating types of reports: ordinary industry reports, in-depth industry reports, and ordinary company reports. YAYI-Ultra accurately counted the number of different types of reports.

Prompt: What is the quantity of each type of report?

When it comes to irregular tables, YAYI-Ultra can still accurately parse and extract key data. The following table contains a total score structure and complex data expressions, and YAYI-Ultra can accurately understand the model types, methods, and locality index changes in the table and complete comparative analysis.

Prompt: Which base model experienced the most significant decrease in locality after using the IKE method?

In terms of statistical data understanding, it can be seen that YAYI-Ultra has significantly enhanced its capabilities in complex layout understanding and cross-language Q&A in table Q&A.

From financial reports and academic papers to complex tables with nested structures, YAYI-Ultra can accurately locate information and understand user intent; at the same time, the model can provide efficient and clear answers in cross-language table Q&A scenarios.

03 Function Call: Intelligent Planning for Complex Tasks

To increase the difficulty, we asked YAYI-Ultra to draw a line chart of the number of gold, silver, and bronze medals won by the Chinese team at last year's Olympics (over time).

First, it can be seen that YAYI-Ultra accurately understood the user's intent, determining that "last year's Olympics" refers to the Paris Olympics, and developed a detailed task plan. Next, the model obtained data related to the gold, silver, and bronze medals won by the Chinese team at the Paris Olympics (including the types of 91 medals and the times they were won) through a search engine; it then organized this medal data, categorized and sorted it by time, and generated code to complete the line chart drawing through the code interpreter.

YAYI-Ultra's ability to complete this series of complex task breakdowns and planning is attributed to its enhanced tool invocation capabilities, which mainly include basic tools such as search engines, code interpreters, image parsing, and weather; as well as specialized vertical tools like news hot list tracking and influence analysis.

The model has significantly improved the rationality of planning in scenarios involving serial calls to multiple tools, while also enhancing its information collection capabilities in complex search scenarios.

04 Multimodal Output: Combining Text and Images, Intuitive and Concise

In literature reading or information collection processes, we often need to search for and analyze specific information (such as numerical changes, experimental results, etc.) from multiple documents. Now, a single sentence can find the desired content, and YAYI-Ultra can simultaneously provide corresponding image content based on text analysis.

For example, asking: The percentage of different behaviors under different collaborative strategies.

YAYI-Ultra identifies relevant multiple AI papers from the "AI Paper Knowledge Base" constructed by the user based on the question and answers accordingly. The answer not only includes text but also provides the original images at the corresponding citation locations, greatly enhancing the reading experience and answer reliability.

05 Full-Stack Long Text: Producing Thousands of Words with Ease

The most impressive feature is the ultra-long text output. YAYI-Ultra supports input of up to 200,000 words and output of 100,000 words, forming a closed loop of full-chain long text capability from "input understanding" to "content creation."

YAYI-Ultra supports both online intelligent creation and literature anchoring creation modes, breaking down long text writing tasks into smaller, more controllable sub-tasks (first generating an outline, then generating the full text based on the outline), effectively ensuring text structure and improving long text generation quality.

● Online Intelligent Creation: Collecting Information Online to Complete Creation

Prompt: Write a 30,000-word analysis report on the development history of Chinese Confucian culture.

●Literature Anchoring Creation: Defining Knowledge Boundaries for Precise Writing

Prompt: Please write a long article based on the reference materials, with the theme "General Artificial Intelligence Solutions: The Perfect Combination of Innovation and Efficiency."

06 Data Analysis: Accurate Solutions, Visual Interaction

Finally, we also conducted tests on basic data analysis and visual chart drawing, where YAYI-Ultra accurately completed analysis, calculation, and chart drawing tasks.

Prompt: Based on the table, calculate the per capita monthly income, then calculate the difference between the monthly income and the per capita monthly income, and draw a bar chart with names on the horizontal axis, the difference on the vertical axis, and the title "Income Difference from Average."

YAYI-Ultra generated and executed Python code through its Python of Thought (POT) capability as per user requirements, accurately completing numerically intensive tasks such as statistical inference, matrix operations, and numerical optimization.

From "Flooding" to "Precise Matching"

YAYI-Ultra with Flexible Expert Configuration

Breaking Through the Bottleneck of Large Model Implementation

Currently, the implementation of AI large models is facing a critical juncture where the "capability-cost" gap is widening.

According to the latest IDC report, enterprises face the issue of model accuracy not fully meeting business needs during the implementation of AI large models; at the same time, 92% of enterprises believe that the lack of computational resources is the biggest challenge during the engineering implementation phase of large models.

The technical team of Zhongke Wenge revealed that YAYI-Ultra is a hybrid expert model characterized by multi-domain capabilities. To enhance performance in specialized tasks across different fields, it adopts a flexible expert configuration model, supporting combinations of experts in various fields such as mathematics, coding, finance, public opinion, traditional Chinese medicine, and security. This can significantly alleviate the common "seesaw" phenomenon in the vertical domain transfer of dense models, providing "high precision, low energy consumption" intelligent solutions tailored to different industry needs.

For example, in the media field, Zhongke Wenge launched the Hongqi 3.0 integrated media intelligent platform, based on YAYI capabilities, helping clients reduce content creation time by 30%-50% and increase content publishing frequency by 20%-40%. After introducing automated review capabilities, one client reduced content error rates from 5% to about 0.5%, and it is now widely used in leading media outlets such as Xinhua News Agency, CCTV, and China Daily.

Zhongke Wenge Hongqi 3.0 Integrated Media Intelligent Platform

In the medical field, the YAYI-based Da Yi Jin Kui traditional Chinese medicine model can accurately diagnose over 500 common diseases and provide personalized treatment plans for patients. Clinical expert evaluations show that the accuracy of syndrome differentiation reasoning is as high as 90%. In simulated tests for the traditional Chinese medicine practitioner qualification exam, it performed excellently with an accuracy rate exceeding 94%, and it has launched the "Da Yi Jin Kui" traditional Chinese medicine health management APP for end users.

China Academy of Chinese Medical Sciences & Zhongke Wenge Da Yi Jin Kui Traditional Chinese Medicine Health Management APP

In the finance and taxation field, the YAYI-based finance and taxation knowledge model achieved an accuracy rate of 90.1% in specialized evaluations, higher than other similar models. After integrating the large model, clients achieved 24/7 uninterrupted consulting services, reducing user wait times by about 50% and increasing user satisfaction by over 30%.

Aerospace Information and Zhongke Wenge Jointly Developed Finance and Taxation Knowledge Model

Currently, YAYI-Ultra (yayi.wenge.com) has opened data analysis, knowledge base literature parsing, and ultra-long text writing functionality for experience on its official website. Interested friends can log in to try it out.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。