In recent days, the "fierce battle" of domestic LLMs heated up again. Ali, Baidu, 360 and other Internet giants have officially announced to upgrade the long text processing function of their AI products, and the explosion of kimi is to blame.

Kimi, also known as Kimi Chat, is a conversational AI smart text assistant built by AI startup Moonshot AI based on its self-developed 100 billion parameter llm, with portals including web pages, apps and WeChat applets.

The software was officially launched in October 2023, the main application scenarios for the translation and understanding of professional academic papers, auxiliary analysis of legal issues, quick understanding of AAPI development documents, etc. It is also the world's first AI text processor due to its 200,000 Chinese characters input.

According to its official website, Moonshot was founded in March 2023, and the core members of the founding team had participated in the research and development of many LLMs such as Google Gemini, Google Bard, Pangu NLP, and Enlightenment. Besides, many core technologies were adopted by mainstream products such as Google PaLM, Meta LLaMa and Stable Diffusion.

It is reported that Kimi has six main functions: long text summary and generation, online search, data processing, code writing, user interaction and translation. In 5 months, the product has once again made a breakthrough in long context window technology.

The data from aicpb.com shows that in February 2024, Kimi received 3.05 million views, with 107.6% increase over the previous month. In domestic AI product list,Kimi ranks third, only after Baidu ERNIE's 10.06 million and Ali Tongyi Qianwen's 3.65 million.

Not only that, Kimi currently has 1 million daily active users and the retention rates continue to rise. As of press time, its app downloads have won the champion in iOS Store (free list of efficiency apps)。

On March 18, 2024, Moonshot announced that Kimi can support 2 million words of ultra-long lossless context, and at the same time release beta version.

According to the current user, Kimi is not an all-round player。The feedback indicates that Kimi performs poorly on conversational continuity for complex tasks, such as structured information processing and scrawled text recognition in images. In terms of generation and voice capabilities, Kimi has not yet fully equipped these modalities.

However, Kimi's long-text abilities speak for themselves。Currently, LLMs on the market like GPT-4Turbo-128k published text range of about 100,000 Chinese characters, and Claude3200k context of about 160,000 Chinese characters.

After the news was pressed, the Kimi concept stock immediately detonated the market. On March 21, Kimi was once down due to a surge in visits. In terms of both the user side and the capital markets, Kimi is experiencing its peak.

Xu Xinran, vice president of Moonshot Engineer, said: "The order of magnitude increase in the length of the lossless context of the large model will further help you to open up your imagination of AI application scenarios, including the analysis and understanding of the complete code base, intelligent agents that can independently complete multi-step complex tasks, life-long assistants that will not forget key information, and multi-modal models with truly unified architectures."

Industry insiders also pointed out that due to the limitations of the model training structure, the text range of LLMs in the past are not long, but complex tasks and industry knowledge is usually very large, in this case, can only split the question before entering, but this may lead to the output of the result up and down incoherent logic is not smooth. Long text makes AI's understanding of context consistent and meets the needs of question and answer.

As a result, Kimi, which is only five months old, once again activated the entire domestic model market. Within 24 hours, large domestic enterprises such as Alibaba, Baidu and 360 stepped in to this AI‘s long text track.

On the evening of March 22, Ali Tongyi Qianwen upgraded and opened the 10 million-word long document processing function for free, becoming the world's first AI application with the largest document processing capacity. From now on, users can quickly read research reports, analyze financial reports, read scientific papers, study and judge cases, read medical reports, interpret legal provisions, summarize in-depth articles, etc. through the Tongyi Qianwen website and App.

On March 23, 360 Wisdom announced the official internal test of 5 million words long text processing function, and will soon enter the 360 AI browser, App version at the same time online.

Next month, Baidu ERNIE will also be upgraded, when it will also open 2 million-5 million words of long text processing functions.

In general, the rapid improvement of the performance of domestic AI models does provide more choices for AI application innovation and lower the application threshold on the one hand. However, on the other hand, downstream applications are also facing homogenization challenges that rely on model providers, and the issue of intellectual property rights to generate content will remain a challenge in the commercialization process.