阿里云Qwen2.5开源大模型发布，性能逼近GPT-4，引领AI领域新高峰
Beijing, China – In a significant development for the artificial intelligence (AI) sector, Alibaba Cloud has unveiled the latest version of its open-source large language model, Qwen2.5. The new model, which was showcased at the Cloud栖 Conference on September 19th, has once again demonstrated Alibaba Cloud’s leadership in the field of AI.
Qwen2.5: A Leap Forward in Open-Source AI Models
Since its initial release in August 2023, the Qwen model has garnered global attention and acclaim from developers. The newly released Qwen2.5 builds upon this success, offering a series of advancements that have positioned it as a formidable competitor in the open-source AI space.
Performance and Capabilities
The Qwen2.5 series is a super AI model ‘treasure box’ that encompasses various sizes of language models, multimodal models, mathematical models, and code models. Each size features a base version, an instruction-following version, and a quantized version, totaling over 100 models. This extensive offering has shattered industry records.
In terms of performance, the Qwen2.5-72B model has surpassed the Llama 3.1-405B, placing it at the top of the global open-source large language model rankings. The Qwen-Max model, on the other hand, has undergone a comprehensive upgrade, with its performance now approaching that of GPT-4o.
Enhanced Capabilities and Applications
The Qwen2.5 models have demonstrated enhanced code, mathematical, and language processing capabilities, along with leading multimodal processing and visual intelligence. This makes them a standout choice in the AI technology landscape.
A Treasure Trove of Models
The Qwen2.5 series includes seven language model sizes: 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B, each achieving the best performance in its respective parameter category. These models have been pre-trained on 18 trillion tokens, showcasing their vast knowledge and strong programming and mathematical abilities.
The Qwen2.5-Coder and Qwen2.5-Math models, specifically designed for programming and mathematics, have also seen substantial improvements. The Qwen2.5-Math model supports the use of思维链 (Thought Chain) and Tool-Integrated Reasoning (TIR) to solve bilingual mathematical problems, making it the most advanced open-source mathematical model series to date.
Multilingual and Multimodal Support
The Qwen2.5 models support a maximum context length of 128K and can generate up to 8K of content. They also possess strong multilingual capabilities, supporting 29 languages, including Chinese, English, French, Spanish, Russian, Japanese, Vietnamese, and Arabic.
In the multimodal model domain, the highly anticipated visual language model Qwen2-VL-72B has been officially released. Qwen2-VL boasts powerful visual understanding capabilities, capable of recognizing images of different resolutions and aspect ratios, as well as understanding video content lasting over 20 minutes. It also supports visual intelligence functions for autonomous operation of smartphones and robots, showcasing its versatile application scenarios.
Qwen-Max: A New Benchmark in AI
In addition to Qwen2.5, Alibaba Cloud has introduced Qwen-Max, a comprehensive upgrade to the backend models of the company’s official website and app. Qwen-Max has been trained with more data, a larger model scale, and deeper human alignment, leading to a significant leap in intelligence levels.
Performance Improvements
In various authoritative benchmark tests, including MMLU-Pro, MATH, GSM8K, MBPP, MultiPL-E, and LiveCodeBench, Qwen-Max has approached the performance of GPT-4o, with particularly impressive gains in mathematical and coding abilities. Compared to the initial version of Tongyi Qianwen in April 2023, Qwen-Max has seen improvements of 46% in understanding ability, 75% in mathematical ability, 102% in coding ability, 35% in anti-hallucination ability, and 105% in instruction-following ability. Moreover, the alignment level of the model with human preferences has seen a qualitative leap, increasing by over 700%.
Conclusion
The release of Qwen2.5 and Qwen-Max marks another major milestone for Alibaba Cloud in the AI sector. With their enhanced capabilities and performance, these models are poised to drive innovation and advancement in the field, solidifying Alibaba Cloud’s position as a leader in open-source AI.
Views: 0