On May 21st, the reporter learned from Alibaba Cloud that the API input price of the Qwen Long, the main model of the Tongyi Qianwen GPT-4, has decreased from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a direct decrease of 97%. This means that 1 yuan can buy 2 million tokens, which is equivalent to the amount of text in 5 New China Dictionary books. This model supports up to 10 million tokens of long text input, and after a price reduction, it is approximately 1/400 of the GPT-4 price.
At the Wuhan AI Leaders Summit held on the 21st, Liu Weiguang, Senior Vice President of Alibaba Cloud Intelligent Group and President of Public Cloud Business Unit, said, "As China's largest cloud computing company, Alibaba Cloud has significantly reduced the price of large model inference this time in order to accelerate the explosion of AI applications. We expect the number of calls to large model APIs to increase by thousands of times in the future."
Liu Weiguang described the new changes in Alibaba's Tongyi Qianwen as "breaking through the global bottom price and accelerating the AI outbreak".
The price reduction covers a total of 9 commercial and open-source series models
It is reported that the price reduction of Tongyi Qianwen this time covers a total of 9 commercial and open source series models, including Qwen Long, Qwen Max, Qwen 1.5-72B, etc. Among them, the main model of Tongyi Qianwen is Qwen Long, with a maximum context length of tens of millions. The API input price has decreased from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a decrease of 97%; The flagship model Qwen Max, just released, has caught up with GPT-4 Turbo in terms of performance on the authoritative benchmark OpenCompass. Its API input price has dropped to 0.04 yuan/thousand tokens, a decrease of 67%.
Among them, the main model Qwen Long performs against GPT-4, can handle ultra long contextual scenarios, supports input in different languages such as Chinese and English, and supports ultra long contextual conversations up to 10 million tokens (approximately 15 million words or 15000 pages of documents). The document service launched synchronously with the Alibaba Cloud Bailian platform supports parsing and dialogue in various document formats such as Word, PDF, Markdown, EPUB, and Mobi.
Public cloud+API will become the mainstream way for enterprises to use large models
As the performance of large models gradually improves, AI application innovation is entering a period of intensive exploration, but high inference costs remain a key factor restricting the large-scale application of large models.
Unlike private deployment, cloud based invocation provides greater space for cost reduction and efficiency enhancement of large models. In general, private deployment of open-source models requires self built clusters, taking into account multiple cost factors such as hardware procurement, software deployment, network costs, electricity costs, hardware depreciation, and manpower. If computing resources are idle or overloaded, additional costs need to be paid; Calling the big model API on the cloud truly achieves on-demand and on-demand use.
Liu Weiguang described the new changes in Alibaba's Tongyi Qianwen as "breaking through the global bottom price and accelerating the AI outbreak".
He stated that whether it is an open source model or a commercial model, public cloud+API will become the mainstream way for enterprises to use large models, mainly for three reasons:
Firstly, the technological dividends and economies of scale of public clouds bring enormous cost and performance advantages. Alibaba Cloud can continuously optimize from both the model itself and the AI infrastructure, pursuing the ultimate inference cost and performance. Alibaba Cloud has built an extremely elastic AI computing power scheduling system based on self-developed core technologies and products such as heterogeneous chip interconnection, high-performance network HPN7.0, high-performance storage CPFS, and artificial intelligence platform PAI. Combined with the Bailian distributed inference acceleration engine, it significantly reduces model inference costs and accelerates model inference speed.
That is to say, even for the same open-source model, the call price on public clouds is far lower than that of private deployment. Taking the Qwen-72B open-source model and a monthly usage of 100 million tokens as an example, directly calling the API on Alibaba Cloud Bailian only costs 600 yuan per month, and the average monthly cost of private deployment exceeds 10000 yuan.
The second is that the cloud is more convenient for multiple model calls and provides enterprise level data security protection. Alibaba Cloud can provide a dedicated VPC environment for each enterprise, achieving computation isolation, storage isolation, network isolation, and data encryption, fully ensuring data security. At present, Alibaba Cloud has led or deeply participated in the formulation of more than ten international and domestic technical standards related to large model security.
The third is the natural openness of cloud vendors, which can provide developers with the richest models and toolchains. On the Alibaba Cloud Bailian platform, hundreds of high-quality models from both domestic and international markets, such as Tongyi, Baichuan, ChatGLM, and Llama series, are gathered. The platform is equipped with a large model customization and application development toolchain, allowing developers to easily test and compare different models, develop exclusive large models, and easily build RAG and other applications. From selecting models, adjusting models, building applications, to providing external services, it's a one-stop solution.
According to the latest data, the Tongyi Big Model has served over 90000 enterprises through Alibaba Cloud and over 2.2 million enterprises through DingTalk services. It has been applied in fields such as PC, mobile phones, automobiles, aviation, astronomy, mining, education, healthcare, catering, gaming, and cultural tourism.
On May 9th, Xiaomi's artificial intelligence assistant "Xiaoai Classmate" reached a cooperation agreement with Alibaba Cloud Tongyi Big Model to strengthen its multimodal AI generation capabilities in image generation, image understanding, and other aspects, and has been implemented on various types of devices such as Xiaomi cars and mobile phones. In addition, companies such as Weibo, Zhongan Insurance, and Perfect World Games have also announced the integration of the Tongyi Big Model and its application in social media, insurance, gaming, and other fields.