DeepSeek Sparks Low-Cost Revolution: How Does This Domestic Large Model Achieve Both High Precision and Low Power Consumption?
TechFlow Selected TechFlow Selected
DeepSeek Sparks Low-Cost Revolution: How Does This Domestic Large Model Achieve Both High Precision and Low Power Consumption?
When concerns about computational power bottlenecks are alleviated, what issues should large models that balance energy consumption and accuracy address?
Navigating Web3 tides with focused insightsSource: Guangzhui Intelligence
Image source: Generated by Wujie AI
In early 2025, DeepSeek triggered a seismic shift in the domestic and global large model industry. Beyond the outstanding performance of its deep reasoning model DeepSeek-R1 in answering questions, DeepSeek's emergence has injected a tense yet vibrant energy into China's large model ecosystem.
First, leveraging its technical advantages, DeepSeek rapidly entered the top tier of international large models, showing domestic companies the possibility of overtaking competitors through a shortcut.
Second, DeepSeek’s training achievements have overcome limitations imposed by computing power constraints, proving that high-quality models can be trained efficiently even with limited computational resources through algorithmic optimization.
As concerns over computing power bottlenecks begin to ease, what challenges should large models balancing efficiency and accuracy focus on? On this front, Chinese large model companies are now presenting their respective solutions.
Recently, YAYI-Ultra, the flagship version of the YaYi large model released by Zhongke Wenge—a Beijing-based AI company incubated by the Institute of Automation, Chinese Academy of Sciences—offered its own answer to resolving the "accuracy-efficiency" dilemma in large model deployment.
As an authoritative evaluation system covering over 100 global models, the OpenCompass leaderboard has long served as a "barometer" for tracking large model technology trends. In its latest public academic ranking, Zhongke Wenge’s YAYI-Ultra scored 64.5 points, entering the global top 10 for the first time and becoming one of only five Chinese large models in the elite group.
In the latest OpenCompass Large Language Model Public Academic Real-Time Ranking, YAYI-Ultra achieved a comprehensive score of 64.5, placing tenth overall, with the following highlights:
Code generation: Ranked fifth on LiveCodeBench, outperforming the GPT-4o-20241120 version
Complex instruction understanding: Ninth place on IFEval
Knowledge reasoning capability: Ninth place on MMLU-Pro
In the C-Eval benchmark, which emphasizes Chinese language understanding, YAYI-Ultra ranked second on the publicly accessible leaderboard allowing autonomous verification, demonstrating strong technical advantages in Chinese-language scenarios.
Hands-on Testing: Ultra-Long Text Output
Precise Handling of Complex Task Planning
According to official information, YAYI-Ultra excels in chart comprehension, complex tasks, and long-text understanding and generation. We immediately tested its performance across 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 Upgraded Visual Understanding: Fluent in Language, Master of Charts
Let’s start with interpreting a chart.
prompt: Around the year 2000, which price range saw the largest change in property management fees?
YAYI-Ultra accurately identifies different colors and numbers in the bar chart, fully understands the visual data, and provides a correct response.
Beyond Chinese contexts, YAYI-Ultra also accurately interprets and follows user instructions in multilingual scenarios, delivering precise 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 is evident that in visual understanding, YAYI-Ultra has comprehensively upgraded its capabilities in addressing technical challenges such as cross-language multimodal alignment, multi-chart reasoning, and variable resolution. The model enhances performance in cross-language chart understanding, multi-chart question answering, and multimodal instruction following. It not only handles complex chart types like stacked bar charts, scatter plots, and hybrid charts with ease but also excels at chart redrawing and transformation tasks.
02 Smart Table Interpretation: No Table Too Complex
In real-world work, compiling statistics from complex reports can be time-consuming. We provided YAYI-Ultra with a table containing alternating entries of three report types—industry general reports, industry deep-dive reports, and company general reports—and the model accurately counted the number of each type.
prompt: How many reports are there of each type?
Even when dealing with irregular tables, YAYI-Ultra precisely parses and extracts key data. For example, given a table with hierarchical structures and complex data expressions, it accurately understands model types, methods, and changes in locality metrics, completing comparative analysis successfully.
prompt: Which base model experienced the greatest drop in locality after applying the IKE method?
In terms of statistical data comprehension, YAYI-Ultra shows significant improvements in understanding complex layouts and supporting cross-language table queries.
Whether financial reports, academic papers, or nested complex tables, YAYI-Ultra accurately locates information and interprets user intent. Additionally, it delivers efficient and clear responses in cross-language table question-answering scenarios.
03 Function Call: Intelligent Planning for Complex Tasks
Raising the difficulty, we asked YAYI-Ultra to plot a line graph showing the number of gold, silver, and bronze medals won by China's team over time during last year's Olympics.
First, YAYI-Ultra correctly interpreted the user’s intent, identifying "last year's Olympics" as the Paris Olympics, and formulated a detailed task plan. Next, it used a search engine to retrieve relevant data on China's gold, silver, and bronze medals (including 91 total medals and their award times). It then organized and grouped the data chronologically, generated code, and invoked a code interpreter to complete the line graph drawing.
YAYI-Ultra’s ability to decompose and plan such complex tasks stems from its enhanced tool-calling capabilities, including basic tools such as search engines, code interpreters, image analyzers, and weather services, as well as specialized vertical tools like news trend tracking and influence analysis.
The model significantly improves planning rationality in multi-tool sequential invocation scenarios and strengthens information gathering in complex search environments.
04 Multimodal Output: Clear, Intuitive, and Illustrative
During literature review or information gathering, users often need to locate and analyze specific information (e.g., numerical changes, experimental results) across multiple documents. Now, with a single query, desired content can be found instantly. Moreover, beyond textual descriptions, YAYI-Ultra can simultaneously present corresponding images.
For example, asking: The percentage of different behaviors under different collaborative strategies
YAYI-Ultra identifies relevant AI research papers from the user’s constructed “AI paper knowledge base” and responds accordingly. The answer includes both text and original figures placed at appropriate citation points, greatly enhancing readability and reliability.
05 Full-Stack Long-Text Generation: Effortless Mastery of Lengthy Content
The most impressive feature is ultra-long text output. YAYI-Ultra supports up to 200,000 characters in input and 100,000 characters in output, establishing a closed-loop capability chain from "input comprehension" to "content creation" for long-form texts.
YAYI-Ultra supports two modes—internet-connected intelligent creation and reference-anchored writing—and breaks down long-form writing tasks into smaller, manageable subtasks (first generating an outline, then expanding it into a full article), effectively ensuring structural coherence and improving output quality.
● Internet-connected intelligent creation: Gather information online to support content creation
prompt: Write a 30,000-character analytical report on the historical development of Confucian culture in China
● Reference-anchored creation: Precise writing within defined knowledge boundaries
prompt: Based on the provided references, write a long-form article on the theme “General Artificial Intelligence Solutions: The Perfect Combination of Innovation and Efficiency”
06 Data Analysis: Accurate Computation with Visual Interaction
Finally, we conducted hands-on testing on basic data analysis and visualization tasks. YAYI-Ultra accurately completed analysis, calculations, and chart generation.
prompt: Based on the table, calculate the average monthly income, then compute the difference between each person's monthly income and the average, and draw a bar chart with names on the x-axis, differences on the y-axis, and title “Difference Between Individual Income and Average”.
Following user instructions, YAYI-Ultra utilized its Python-of-Thought (POT) capability to generate and execute Python code, accurately completing numerically intensive tasks such as statistical inference, matrix operations, and numerical optimization.
From "Flood Irrigation" to "Precision Matching"
YAYI-Ultra Breaks Through Large Model Deployment Bottlenecks
with Flexible Expert Configuration
Currently, large AI model deployment faces a critical juncture where the "capability-cost" gap is widening.
According to the latest IDC report, enterprises face challenges in deploying large AI models, particularly in meeting business needs due to insufficient model accuracy. Meanwhile, 92% of enterprises identify lack of computing resources as the biggest challenge during the engineering deployment phase of large models.
Zhongke Wenge’s YaYi technical team revealed that YAYI-Ultra is a mixture-of-experts model designed for multi-domain capabilities. To enhance performance on specialized domain tasks, it adopts a flexible expert configuration mode, supporting combinations of experts in mathematics, coding, finance, public opinion, traditional Chinese medicine, security, and more. This approach significantly mitigates the common "see-saw" effect encountered when dense models are adapted to vertical domains, delivering "high-precision, low-energy-consumption" intelligent solutions tailored to diverse industry needs.
For instance, in the media sector, Zhongke Wenge launched the Red Flag 3.0 intelligent media convergence platform. Built on YAYI, it helps clients reduce content creation time by 30%-50%, increase content publishing frequency by 20%-40%, and lower content error rates from 5% to around 0.5% after introducing automated review capabilities. It is now widely adopted by leading media outlets such as Xinhua News Agency, CCTV, and China Daily.
Zhongke Wenge Red Flag 3.0 Converged Media Intelligent Platform
In healthcare, the Da Yi Jin Kui TCM large model, powered by YAYI, can accurately diagnose over 500 common conditions and provide personalized treatment plans. Clinical expert evaluations show diagnostic reasoning accuracy reaching 90%, with over 94% accuracy in simulated TCM practitioner qualification exams. A consumer-facing app, “Da Yi Jin Kui,” has been launched for TCM health management.
China Academy of Chinese Medical Sciences & Zhongke Wenge – Da Yi Jin Kui TCM Health Management App
In finance and taxation, the YAYI-powered tax and finance knowledge large model achieved a 90.1% accuracy rate in specialized assessments—outperforming other similar models. After integrating the large model, clients offer 24/7 consulting services, reducing customer wait times by about 50% and increasing user satisfaction by over 30%.
Aerospace Information & Zhongke Wenge Jointly Developed Tax and Finance Knowledge Large Model
Currently, YAYI-Ultra (yayi.wenge.com) offers free trial access on its official website for features including data analysis, knowledge base document parsing, and ultra-long text writing. Interested users are welcome to log in and try it out.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
