In the AI competition, the “battle” of overseas giants is intense, and the evolution of big domestic models is also accelerating.
Held in Beijing on May 9$BABA-SW (09988.HK)$At the Alibaba Cloud AI Leaders Summit, Tongyi Qianwen 2.5 was officially released.
This time, the big model not only achieved major upgrades in capabilities, but also completely overtook GPT-4 Turbo in performance, making it the “strongest on the surface” Chinese model.
Overtake GPT-4
Alibaba Cloud said that compared to Tongyi Qianwen 2.1, Tongyi Qianwen 2.5's ability to understand, reason, follow instructions, and code was increased by 9%, 16%, 19%, and 10%, respectively.
Compared to GPT-4, in the Chinese context, Tongyi Thousand Questions 2.5 comprehensively surpasses many abilities such as text comprehension, text generation, knowledge questions and answers & life suggestions, small talk and conversation, and safety risks.
In this wave of AI big models motivated by competition, this is the first time that a large domestic model has achieved such results.
In terms of long text processing, Generic Thousand Questions 2.5 supports processing of documents up to 10 million words at a time, and can process up to 100 documents at the same time.
Tongyi Qianwen also supports the processing of documents in various formats such as PDF, Word, and charts, to meet the diverse needs of users.
In addition to Tongyi Qianwen 2.5, Alibaba Cloud also released a new set of “records.”
Alibaba Cloud Chief Technology Officer (CTO) Zhou Jingren said at the conference that the Tongyi Big Model has passed more than 90,000 Alibaba Cloud service companies and more than 2.2 million through DingTalk.
The number of daily calls to the Tongyi Qianwen API has exceeded 100 million, and the cumulative number of downloads of the Tongyi Open Source Model has exceeded 7 million.
In addition, the process of implementing Tongyi is also accelerating. Currently, it is involved in various fields such as PCs, mobile phones, automobiles, aviation, astronomy, mining, education, medical care, catering, games, and cultural tourism.
Among them, Xiaomi's “Little Ai Classmate” has also reached a cooperation with Alibaba Cloud's Tongyi Big Model and will launch various devices such as Xiaomi cars and mobile phones.
In addition, Tongyi Qianwen's latest open source 110 billion parameter model, Qwen1.5-110B, also achieved the best results.
In benchmark tests such as MMLU, TheoremQA, and GPQA, this model surpassed Meta's Llama-3-70B to become the most powerful model in the open source field.
Tongyi Spirit Code announced the launch of the Enterprise Edition. Tongyi Spirit Code is the intelligent coding assistant with the largest number of users in China. It was developed based on the SOTA-level Tongyi Thousand Question Code model CodeQwen1.5, and the number of plug-in downloads has exceeded 3.5 million.
At the summit, Alibaba Cloud emphasized becoming “the most open cloud in the AI era” to help customers seize the opportunities of the big model era through open computing power platforms, open source self-developed models, and high-quality model services.
Is the big domestic model entering the core arena?
In April of last year, Tongyi Qianwen officially launched.
Coinciding with the first anniversary, the advancement of the Tongyi Thousand Questions model also meant that the big domestic model was taken to the next level.
Since the launch of ChatGPT in 2022, the big AI model has set off the largest wave of artificial intelligence in history around the world.
It can be said that in the past year, this circle “needed more volumes, more volumes.”
Currently, OpenAI, Google, Microsoft, etc. are aggressively “throwing money” and continuously innovating their products. In addition to the overseas giants “You Catch Up with Me,” big domestic models are also “doing their best” to follow suit.
According to research data from the SuperClue team, the progress of the domestic big model is roughly divided into three stages, namely preparation period, growth period, and explosion period.
Currently, in addition to technology giants such as Ali Tongyi Qianwen, Baidu Wenxinyi, and Huawei Pangu to step up the development of major domestic models, Fudan MOSS, Zidong Taichu of the Chinese Academy of Sciences, and Kimi, the startup Dark Horse - Dark Side of the Moon, stand out one after another.
Judging from the development of the industry, Galaxy Securities said that in the future, the field of general AI models may be more concentrated on leading manufacturers. More manufacturers need to specialize in the industry. Vertical AI models and end-side AI models will be the main battlefields in the future, and the market space is broad.
It is recommended to focus on opportunities related to upstream computing power infrastructure, opportunities related to domestic computing power industry chains and ecological partners, and downstream application areas.
Editor/Jeffrey