Quark App Releases Self-developed Large-scale Model with Billion-level Parameters, Breaking the Record of Chinese Language Understanding in CMMLU

2023-11-14

The Quark Mega Model is a fully self-developed, multimodal model with billions of parameters launched today by Alibaba's Intelligent Information Business Group. It will first be applied to various apps under Quark, transforming them into "AI assistants" that provide users with general knowledge Q&A and professional search services. Based on the Transformer architecture, the Quark Mega Model has a comprehensive Chinese database and undergoes daily training and fine-tuning on billions of textual and visual data. It boasts low cost, high responsiveness, and strong overall capabilities. In the CMMLU Mega Model Performance Evaluation, the Quark Mega Model ranks first, surpassing the overall capabilities of GPT-3.5 and even outperforming GPT-4 in certain scenarios such as writing and exams. The Quark Mega Model not only derives models for general knowledge, healthcare, education, and other verticals, but also provides professional services such as AIGC and intelligent retrieval. It has achieved remarkable results in domestic professional exams, with near-perfect scores in the college entrance examination and a passing score of 486 in the clinical practitioner qualification exam. Additionally, it can identify, answer, and guide against misinformation and false information. In the future, the Quark Mega Model will be applied to search, intelligent tools, asset management assistants, and other scenarios. A series of AI-native applications will also be launched.