YAYI-Ultra - A hybrid expert model launched by Zhongke Wenge
YAYI-Ultra is the flagship version of the enterprise-level large language model developed by Zhongke Wenge, possessing powerful multi-domain professional capabilities and multimodal content generation capabilities. It supports multiple fields such as mathematics, coding, finance, public opinion, traditional Chinese medicine, and security...
What is YAYI-Ultra?
YAYI-Ultra is the flagship version of Zhongke Wenge's enterprise-level large language model, possessing powerful multi-domain professional capabilities and multimodal content generation capabilities. It supports expert combinations from multiple fields such as mathematics, coding, finance, public opinion, traditional Chinese medicine, and security, mitigating the "seesaw" phenomenon in vertical domain migration. YAYI-Ultra supports inputs up to 128k and longer context windows, and has multimodal capabilities for aligning over 10 million image and text data. It supports multi-turn dialogue role-playing, content security risk control, and the invocation of 10+ intelligent plugins.
Main functions of YAYI-Ultra
- Multi-field professional capabilitiesYAYI-Ultra is a hybrid expert model that supports combinations of experts from multiple fields such as mathematics, coding, finance, public opinion, traditional Chinese medicine, and security, alleviating the "seesaw" phenomenon that is common in the migration of dense models to vertical fields.
- Multimodal content generationIt supports outputting text and images in a visually appealing manner, and can extract relevant information from the knowledge base based on the user's question, simultaneously providing the corresponding image content.
- Long article outputIt supports input of up to 200,000 characters and output of up to 100,000 characters, forming a closed loop of long text capabilities across the entire chain from "input comprehension" to "content creation".
- Data Analysis and VisualizationIt can accurately complete data analysis, calculation, and chart drawing tasks.
- Intelligent planning for complex tasksYAYI-Ultra enhances the planning rationality in scenarios involving serial calls of multiple tools. It can break down complex tasks into multiple sub-tasks and complete the tasks by calling basic tools such as search engines, code interpreters, and image parsing tools, as well as vertical tools such as news hot list tracking and dissemination influence analysis.
- Internet-connected intelligent creationIt supports collecting information online to complete creative work, such as writing an analytical report on the development history of Chinese Confucian culture.
The technical principle of YAYI-Ultra
- Hybrid expert architectureYAYI-Ultra employs a hybrid expert model architecture, supporting combinations of experts from various domains, such as mathematics, coding, finance, public opinion, traditional Chinese medicine, and security. It significantly alleviates the "seesaw" phenomenon commonly encountered when migrating dense models across vertical domains, providing high-precision, low-energy intelligent solutions for different fields.
- Instruction fine-tuning and data-drivenYAYI-Ultra is trained on a large-scale, high-quality multilingual corpus and combined with instruction fine-tuning techniques, resulting in outstanding performance across different domains and tasks.
YAYI-Ultra's performance review
In the OpenCompass large model public academic leaderboard, YAYI-Ultra entered the top ten for the first time with a score of 64.5, becoming one of the five Chinese large models in the TOP 10. Its specific performance is as follows:
- Code generationIt ranks fifth in LiveCodeBench, outperforming version GPT-4o-20241120.
- Understanding complex instructionsIt ranks ninth in IFEval.
- Knowledge reasoning abilityIt ranks ninth in MMLU-Pro.
YAYI-Ultra's project address
- Github repository:https://github.com/wenge-research/YAYI2
- HuggingFace model library:https://huggingface.co/wenge-research
Application scenarios of YAYI-Ultra
- Media fieldWe help clients reduce content creation time by 30%-50%, increase content publishing frequency by 20%-40%, and reduce content error rate from 5% to around 0.5%.
- medical fieldBased on YAYI's "Da Yi Jin Gui" TCM model, it can accurately diagnose more than 500 common diseases with a diagnostic reasoning accuracy rate of up to 90%, and has launched the "Da Yi Jin Gui" TCM health management APP for C-end users.
- Finance and taxation fieldBased on YAYI's comprehensive financial and tax knowledge model, the model achieves an accuracy rate of 90.1%, higher than other similar models, providing clients with 24/7 uninterrupted consultation services.