Title: Dingjiao Original
Author: Dawn
Editor: Wei Jia
A startup that's only been around for a year, led by a group of post-90s graduates from prestigious universities, has created an outstanding large-scale product, shaking up the entire AI community.
Since the brief fame of the Miya Camera, there hasn't been a blockbuster product in China's large-scale model industry for a long time. Kimi fills this gap.
What's most surprising is that even A-share investors are jumping on the bandwagon, with major stock trading groups discussing Kimi concept stocks, causing a surge in traffic that crashed Kimi's servers.
Now, over a month has passed, and although the hype has subsided, people are still discussing Kimi.
A college student at a university in Beijing told Dingjiao that the school recently organized a study session, arranging for teachers to learn about Kimi, with assignments given to students to learn and train with Kimi.
Online, from time to time, you can see people recommending Kimi. Some of these may be targeted advertising tweets, but there are also genuine recommendations.
One cognitive shock that Kimi brings is how rapidly AI is developing, with startup companies able to produce impressive products in a short time. In comparison, those tech giants that claim to have been working for over a decade and have poured billions of dollars into research seem to lag behind.
Kimi has stolen the limelight from the tech giants, and in the future, it may even steal their users. The question is, can this frenzy continue?
Kimi is an intelligent assistant aimed at end users, proficient in reading long texts, and searching the web. Its main features include organizing information, interpreting documents, assisting in programming, and copywriting, earning it the nickname "ChatGPT Chinese version" among netizens.
With ChatGPT already established in the education market, and products like Wenxin Yiyuan, Tongyi Qianwen, and Xunfei Xinghuo from big companies competing internally, many people were already familiar with the functionalities Kimi offers.
However, unlike others, Kimi is not just self-promotion from the manufacturer; it's a domestically produced large-scale model product that has excited 200 million A-share investors.
What Kimi presents to users is simple: a search box similar to Baidu's, "drag in files; send out URLs," and then users can directly ask it questions regarding the files or URLs sent to it, allowing it to help summarize, analyze data, and search the entire web.
The highlight of Kimi.ai is its support for ultra-long text input. Initially starting with 200,000 characters, it later increased to 2 million characters, surpassing any other large models globally at the time, none of which could exceed 100,000 characters.
The most significant change in user experience is that with Kimi, you no longer need to split your files into multiple parts when feeding them to large models. You can simply throw them in, and Kimi quickly understands and can even pinpoint information within, telling you on which page certain information appears.
With this feature, Kimi's purpose becomes clear: long-text reading and summarization analysis, akin to AI search + document summarization. It serves as a productivity tool to handle information, more practical than for artistic or entertainment purposes.
This positioning determines that Kimi's main users are primarily working professionals. Kimi's official documentation mentions six categories of users: academic researchers, internet professionals, programmers, self-media and content creators, financial and consulting analysts, and legal professionals. They share a commonality: the need to process large amounts of document information.
A non-litigator lawyer told "Dingjiao" that they used to use ChatGPT to organize legal policies, write reports, and summaries. However, ChatGPT had a limitation that each input couldn't be too long. For example, a 10,000-word document needed to be split into several parts and fed to it for analysis. Kimi's long-text feature solves this problem. Now, they use both products simultaneously.
A self-media practitioner told "Dingjiao" that they use large models to analyze financial information of listed companies and query some financial data. After comparing Wenxin Yiyuan and Kimi, they found Kimi to be more user-friendly. "Kimi's features are direct, simple to use, and it has strong summarization capabilities."
However, they all indicated that the current usage frequency is not high, mostly for trial and exploration, because the quality of AI-generated content is unstable. Sometimes, it may produce nonsense. "It can be referenced, but not fully trusted."
Kimi's developer is a startup called "Moon's Dark Side," established in March last year. Kimi made its first appearance in October last year, initially focusing on long-text processing. Its marketing gimmick was claiming to have "fully digested a copy of 'The Three-Body Problem.'"
Within the six months after its debut, Kimi iterated three times, adding features such as recognizing scanned documents, launching a mini-program, enabling online search, and upgrading the supported context length to 2 million characters.
The 2 million characters upgrade on March 18 was a turning point. Before that, Kimi was mainly circulating among internet circles, large model practitioners, and a small group of AI enthusiasts. After the upgrade, Kimi caught the attention of 200 million stock investors, leading to a surge in Kimi concept stocks, and Kimi APP and mini-program crashing and trending on social media, further spreading its popularity.
Looking at it now, Kimi's success is the result of several factors. Firstly, the product is indeed good, which is a prerequisite. Secondly, timely and effective marketing, such as Moon's Dark Side's $1 billion financing in mid-February, claiming to be the largest single-round financing for a domestic AI large model company, greatly increased its visibility. In addition, the fermentation of Kimi concept stocks, with the help of 200 million stock investors, further propelled Kimi onto Weibo's hot search and the top ten of the App Store's overall rankings.
Kimi has indeed gained traction, unlike previous domestic large models that were only popular within industry circles. However, in terms of overall usage, the actual number of people using Kimi is still limited. Similarweb monitors show that the peak daily active users of Kimi's web version are around 300-400 thousand, with overall daily active users reaching millions.
For a startup company, this achievement is quite impressive. However, the question remains: why was it a startup company that achieved this?
Moon's Dark Side was not among the earliest companies in China to develop large models. Before it, Baidu's Wenxin Yiyuan, as the first domestic large model after ChatGPT, was considered the closest to ChatGPT. In addition, Alibaba's Tongyi Qianwen, iFLYTEK's Xinghuo, and Zhipu GLM have all undergone several iterations.
However, despite the manufacturers' active promotion, there has been no product in China that C-end users truly recognize and are willing to promote spontaneously.
One major chaos is parameter tuning and ranking manipulation. Almost every company, when releasing a product, would compare it with GPT and find a few metrics to surpass it—such as Chinese proficiency. GPT became a target, being overtaken by domestic large models one after another. However, industry insiders know that in terms of overall capabilities, there is no domestic large model that can surpass GPT4.
"Ranking manipulation" is an open secret in the large model industry. Domestic large model companies are eager to manipulate rankings and often rank first on various lists. However, many insiders who evaluate rankings told "Dingjiao" that most rankings are not of significant reference value, similar to "cheating" in exams, not representing real capabilities.
Ordinary users are very confused. "You say you are very strong, you are on the rankings, recommended by the media, praised by bloggers, so I believe you, but after using it, it's hard to say..." said a user who has tried many domestic large models.
It's interesting that domestic companies developing large models have various motivations: some are for boosting stock prices, some are riding on hot trends, and some just want to raise money, with very mixed intentions. The better ones aim to surpass OpenAI, aiming for victory.
Even for companies that claim to prioritize long-termism, many are more focused on defeating their competitors rather than satisfying user needs.
The most typical examples are Alibaba and 360. After Kimi announced its support for 2 million characters of ultra-long lossless context, its servers crashed due to a surge in traffic. Then, the next day, Alibaba's Tongyi Qianwen announced free access to processing 10 million characters of long documents, claiming to be the "world's number one in document processing capacity," followed by 360 AI browser announcing the internal testing of 5 million characters of long-text processing functionality. Well, the big players have "won" again, claiming the top spot once more.
Critics are saying, "If you (Alibaba, 360) know long texts are good and you can do it too, why wait? Are you here to provide convenience to users or just to grab traffic?"
Besides the giants, there are also big shots. Last year, shortly after Kimi's release, Zero-to-Infinity, founded by Li Kaifu, unveiled the large model Yi-34B, capable of handling around 400,000 words, about twice the size of Kimi. They claimed it topped several global English and Chinese proficiency test rankings. However, industry insiders quickly pointed out that the architecture of the "Yi series" models, compared to Meta's open-source large model LLaMA, only changed the names of two tensors, sparking controversy over its originality.
Source: Zero One Everything Official Website
In such a vast entrepreneurial ecosystem, with a slightly better product, clever marketing, and good timing, it's quite easy to stand out.
In March this year, AI entrepreneur Hua Rongqi utilized Kimi to write a comprehensive "Kimi Guide" on a cloud document, which gained wide dissemination. He mentioned to "Dingfou" that Kimi has been excelling in long-form text from the beginning, making it a focal point breakthrough, leaving a deep impression. Moreover, founder Yang Zhilin's persona as a super scholar, AI expert, and post-90s figure with a built-in filter, has garnered more external support on an emotional level, and user-initiated publicity has added considerable traffic and goodwill to Kimi.
Product positioning determines market strategy. Among domestic large-scale model products, Kimi is one of the few that has been clear from the outset about being solely focused on the consumer market (to C) and not catering to businesses (to B). On the first day of Kimi's release, Yang Zhilin mentioned the desire to enhance model capabilities while also focusing on super applications for consumers, making Kimi the first product attempt.
The emphasis on consumer users ensures that Kimi's product experience remains top-notch. Whether it's handling 2 million-word-long texts, drag-and-drop file operations, or the sleek page design, all reflect the company's user-centric approach.
Big players like BAT have been in the game early and have diverse aspirations. Regarding the question of "who it's for," big players typically aim for both consumer and business markets, bundling large models with cloud services, trying to cater to both.
Zhifu AI is a startup affiliated with Tsinghua University, directly competing with OpenAI, acknowledged for its robust technical prowess. While Zhifu also has consumer-oriented products, its main focus lies in serving businesses — evident from its large-screen advertisements at airports, aimed at accelerating various industries towards the era of general artificial intelligence.
In comparison, Kimi stands out as one of the few products crafted from the perspective of user needs. This isn't to say that the motivations behind the scenes are necessarily nobler; it's just that, judging by the results, Kimi enjoys higher user acceptance.
As a startup, Moon's Dark Side carries less baggage, garnering more tolerance from people, which is an advantage big players lack.
When Baidu launched Wenxin Yiyuan in China in March last year, people's focus wasn't on the product's functionality but rather on whether it was a rip-off and why the product demonstration was prerecorded.
Consumer users have a natural skepticism towards big players, sometimes even viewing them as enemies of innovation. So, when Yang Zhilin appeared as a post-90s entrepreneur, he was easily labeled as a disruptor and game-changer.
So, what's the real technical strength behind Moon's Dark Side? Can Kimi's popularity last?
Chiguang Yao, founder of AI startup Language Core, believes that overall, considering its ability in handling long-form text, Kimi is a highly promising model and can be ranked among the top tier in China.
His analysis on "Dingfou": "Kimi's ability to trace long texts is impressive, accurately retrieving specific information based on given materials, even buried within tens of thousands of words in a novel. Additionally, its logical capability in handling most daily tasks is sufficient. Although it may miss some details in summarizing and extracting from extremely long materials, it retains the main essence accurately."
Source: Screenshot from kimi.ai official website
Huaronqi believes that each major model has its own strengths: Kimi excels at interpreting long texts, acting like a super-intelligent text retriever; Zhifu's intelligent entity is more comprehensive, establishing an open-source ecosystem, making project development convenient within this ecosystem; MiniMax's model understands human nature better, with advantages in role-playing and emotional value; Wenxin's security and compliance are relatively well-done.
"Kimi is indeed the most dazzling among consumer tools, but in other aspects, such as AI system construction, relying solely on Kimi is not enough," said Huaronqi.
From this perspective, Kimi only excels in some aspects that are most easily perceived by consumer users, temporarily achieving a better experience, but this does not mean it has a significant advantage.
AI industry investor Liu Di told "定焦" that existing large models lack comparability under the same framework. In terms of comprehensive capabilities, products from Baidu, Alibaba, and Zhifu AI are good. However, Kimi is handy in some office scenarios, such as complex searches, multi-turn conversations, and report analysis, and is used by users as an auxiliary tool.
However, due to its clear product positioning, Kimi has already had a substitutive effect on products from major companies, such as search engines.
"I think with Kimi, you can basically stop using search engines," a Kimi user told "定焦." In his opinion, Kimi's online search + automatic generation function is a dimensional reduction strike against traditional searches. "In the past, when searching for information, you had to sift, judge, and analyze from a large number of results by yourself. Now, these processes can be omitted, and you just need to give instructions."
Many people didn't notice that during the time when Kimi was booming, another product also experienced rapid traffic growth. According to statistics from "AI Product Rankings," the AI search product "Mita AI Search" saw a fivefold increase in traffic in March compared to the same period last year, quickly climbing in the overall rankings, second only to Wenxin and Kimi.
Liu Di believes that Kimi has influenced the strategies of major companies. "Major companies have clearly shifted their focus to C-end fine-tuning, attaching more importance to the quality of paperwork rather than just generating some chat content for users to play with."
Of course, some people are skeptical about Kimi's sustainability, believing that it may not necessarily make money. Domestic large model manufacturers engage in B-end business because they can see clear monetization scenarios. Consumer-end products are difficult to become hits, and after a hit, the cost of computing power is a huge burden, while making users pay tests user stickiness. And many users flock to Kimi because it's free.
Minzhi Technology CEO Min Kerui once bluntly stated: Compared to the international market, the willingness to pay in China, regardless of the environment, is poor.
"Kimi hasn't succeeded either. It is trying to cultivate user habits first. Now it can maintain millions of DAU (daily active users), with monthly reasoning costs reaching tens of millions. Next, everyone will have to verify charging," Liu Di said.
Huaronqi believes that charging for C-end users is an inevitable path for Kimi, "It just depends on which stage and which functions to charge for, which puts a lot of pressure on Kimi's product and pricing level."
After major companies changed their strategies, they formed a siege against Kimi, so the speed of technological and product iteration is crucial.
"In this wave of AI, everyone reached a consensus very quickly. Once it enters the white-hot stage of fighting, whether it's snatching users or income, it probably won't last very long," said Huaronqi.
*Image source: Unsplash. At the request of the interviewee, Liu Di is a pseudonym.
