The big model "six tigers" of 2024: the big factory is strong, difficult to find a way out

Over the past 12 months, almost all of the Big Model Six Tigers have had a "have to choose" moment. Only in the first half of 2024 they were still choosing "which way to go".In the second half of the year, we will have to think about "which way is left to go".(Editor's note: The Six Tigers of Big Models is the collective name for six domestic startups that are leading the way in big model development, including Wisdom Spectrum, MiniMax, Dark Side of the Moon, Baichuan, Zero-One Everything, and Step Star.

An investor who pitched into Six Little Tigers in 2024 told "Four Woods Relativity" that in the first half of 2024, most of the Little Tigers' financing was still in a blaze of glory.

He gave an example to show where they were at the height of their powers: at the time, not only was the star company represented by Kimi not open to all due diligence, but there were plans by the old shareholders to set up an investment mechanism for some of the smaller shareholders who wished to get in on the action, and earn new management fees from it themselves.

But in the second half of the year, in the investment market and the AI industry, the reputation of AI startups, including Six Tigers, began to take a sharp turn for the worse. "There are basically no potential buyers now, except for state capital." The aforementioned investor said.

In addition, "Four Woods Relativity" also learned that there are one or two small tigers hoping to find international capital support in the long term, but did not get the result.

In addition to financing being difficult, business is even more difficult.

In the first half of 2024, the six tigers were still considering whether to do To C or To B, focusing on the domestic market or entering the global market. In the second half of the year, we found that the domestic C-end user loyalty is not high, and the B-end money is very difficult to earn, at least three of the six little tigers have seriously considered going to sea.

However, after Dark Side of the Moon's phased contraction of its overseas business, another star big model company recently laid off its North American team.MiniMax's overseas fist product, Talkie, also disappeared from the U.S. App Store in mid-December.

I can't say that the big model companies haven't improved over the year, but they are far from what the market expected a year ago.

This predicament is not limited to China's Six Tigers.Both aggressive domestic majors and overseas AI companies have had a torrid year.

 Status quo: giant "churning", very few profitable business

First of all, the To C model, represented by the general assistant, is no longer a "paradise" for startups to play.

"The next goal is to equal Doubao's current number of users." A business leader of Head Six Tigers told us that this is the expectation that this company has set for its general assistant product after a recent business review.

As we approach the end of the year, a number of companies are beginning to review their business.

An investor close to Kimi said its original goal for 2024 was 10 million daily activity, and it's now about 1/3 of the way there.

10 million, which happens to be close to the recent daily activity of Beanbag App.

During the beta, Doubao was still called "Grace", and arrived late in September 2023, after Kimi, Wenxin Yiyin, Tongyi Qianqian, Zhipu Qingyin, and even ConchQi (now ConchAI) had been launched.

But a year later, the comprehensive AI product list, quantum bit think tank and other channels of data, bean bag in November 2024, the monthly activity has been close to 60 million, the daily activity of more than 9 million, is the first of the domestic general assistant fault.

Among several entrepreneurs and investor friends we've been talking to closely lately, most believe that the biggest advantage of Doubao App's meteoric rise is its control of traffic channels.

According to mobile marketing platform AppGrowing data, since 2024 (as of November 15), Kimi, Doubao, Starfield and other ten domestic big model products, together placed more than 6.25 million ads, with an amount of 1.5 billion yuan.

The amount of these investments, which have been hotly publicized in various channels, has been questioned by several "protagonists" who questioned the exaggeration of the data.

But at least for the big manufacturers who hold the traffic channels, the amount of placement really does not use so much. The internal business of big factories placing their own channels will often use special settlement standards, "sometimes internal departments will share a portion of the traffic, which can also be exchanged for other internal resources." A big factory AI practitioner told us.

This is a perk that startups can't have.

An AI star startup that received a giant investment in 2024 with a valuation of more than $100 million has told "Four Woods Relativity" that because the field it focuses on and the direction of a large company's heavy bets basically overlap, it is unable to place itself on the traffic channels controlled by that large company.

Kimi is no exception. Not long ago there was a media report that Kimi, who swiped and gained a large number of users on content APPs such as Jitterbug and B Station at the beginning of the year, was restricted from placing on Jitterbug in the second half of the year and had to switch to platforms such as Shutterbug.

According to "AI New List" public statistics, as of October 29, Kimi's total number of advertisements in the past three months has exceeded 2,500 items. Among them, more than 2,100 pieces of content have been placed on the Racer platform, accounting for almost 84% of the total amount of advertisements.

There are also some startups that failed to complete funding when hot money was pouring in, couldn't afford the growing cost of traffic, and had to transform.

The CEO of an AI education company founded in 2024 told us that his own early product was priced at $20, but now it costs at least $25 to place a person, and "we just can't do the math."

Secondly, outside of the To C model, the big players figure to prey on the To B market for big models as well.

Overseas, big modeling companies selling APIs is an okay business. openAI expects to make $500 million off of APIs in 2024, or 15% of total revenue. according to CNBC, Anthropic is even more reliant on APIs, with 60% to 75% of total revenue expected to come from third-party calls to APIs.

In China, the provision of large model API, more and more into the model manufacturers to capture the market way, has not been pinned on the lucrative expectations.

A mid-year wave of price cuts, spearheaded by DeepSeek and Byte, limited the price space for APIs.

In May, DeepSeek launched DeepSeek-V2, the price directly "kill" to every million input tokens 1 yuan, every million output tokens 2 yuan.

Also in May, ByteDance set the inference input price for the beanbag generalized model pro-32k version at $0.0008/thousand tokens, which is 99.31 TP4T below the industry average.

In the face of the price offensive, competitors can only respond quickly. Ali lowered the price of the main model of Tongyi thousand questions by 97%; Baidu announced that the two models of Wenxin were free.

"Shiki Relativity" learned that in early 2024, Byte internally set the goal of becoming the No. 1 in the LLM call market. Among them, the proportion of enterprise customers should be close to half. To this end, the volcano engine also set up a team of dozens of algorithms and services to help customers do consulting, testing Prompt, consume a large amount of "high-quality customers" have the opportunity to get more concessions.

A big model sales of a large factory told us that a customer with a daily Token consumption of more than two billion now spends only a small few thousand per month with the beanbag API. But in other companies, such a model is not necessarily supported.

By the end of the year, Byte had already gained a number of public cloud customers in areas such as retrieval, conversation, entertainment, gaming, and customer service.

Third, the only business model left, there is the private deployment of serving Big B customers.

This field, at present, is basically the territory of state-owned enterprises and Xunfei and Baidu.

Smart Spectrum was, for a long time, the only startup to brush its presence from this. By the second half of 2024, Bacchus and Zero One, too, saw single-digit public wins.

Compared with the deployment price of 10 million in early 2023 for large models, now some pure software, small millions of single, counting labor costs and service cycle, is likely to lead to AI companies "do a single loss of a single". But even so, in the non-bidding channels, the AI company's privatized deployment business has begun to white-hot scramble.

One of the six tigers, which joined the Big B market in the second half of 2024, tried to communicate with its customers at market prices, but then encountered another old "Four Little Dragons" that deliberately lowered its prices, and the customers chose the low-priced party.

As a final point, giants that believe in the public cloud are also beginning to plan for the private market.

Not long ago, Volcano Engine President Tan Toi said that while he's more bullish on using big models on the public cloud, the hybrid model will persist next.

In the second half of 2024, the volcano engine began to appear frequently in bidding information. Byte's recently launched AI application development platform "HiAgent" will also carry Beanbag's privately deployed version as it focuses on enterprise scenarios.

In fact, while many people first heard of HiAgent at the volcano's "Original Power Conference" held at the end of the year - previously better known as Coze/Button, which is more developer-oriented - HiAgent has been a low-key presence for at least half a year.

At that time - mid-2024 - a number of Volcano's internal employees hadn't realized that Byte also wanted to offer some private deployment options. Some AI company employees fed back that in the following months, Byte's sales started promising the public that they could help companies deploy and fine-tune for free.

In the face of an all-out attack by the giants, the direction of the startup's business has become erratic.

In October, news broke that a number of the six tigers had abandoned pre-training. Although, continuing to blindly raise the number of participants may lead to an increase in modeling effectiveness that is not proportional to the investment. But this action will still put a question mark on their valuation of several billion dollars.

Two investors have told us that a Tiger that is not valued in the top three and is set to transition into the pendant space has even sold the cards used as pre-training.

Industry-Wide: Killer Apps Not Available, Technology Desperately Needs a New Paradigm

The contrast is that the already multi-billion dollar drawingAI Universal Assistant, not the Killer App in everyone's mind.

Not long ago "Intelligent Surge" reported that even though Doubao is the number one in the user break, Byte executives believe that it is only an intermediate state of the product, and the priority of the other app, namely Dream, has been elevated.

We observed that Instant Dream is continuously trying to take AI-generated content and turn it into a stream of feeds in the app. This detail refracts that while Instant Dream has a much weaker focus than Beanbag, it is perhaps more AI Native.

More AI Native applications that are still in the process of running out.

In March 2023, the open source project AutoGPT burst into flames, allowing Agents capable of independently completing human tasks to come to the front of more people.At the end of 2023, after the launch of OpenAI's GPTs, large domestic manufacturers began to scramble to launch a series of Agent building platforms such as Coze/Button, Baidu Wenshin Intelligent Body Platform, and so on.

However, from the actual performance, whether it is booking a flight, or buying coffee, Agent is still difficult to be used smoothly. For example, McDonald's AI ordering system, which uses Agent, has been shut down because it often places orders incorrectly. Air Canada's Agent - consulting and claims robot also because of randomly give unauthorized coupons were taken offline.

(McDonald's AI ordering system incorrectly added hundreds of dollars of chicken nuggets to customers)

The root cause of Agent's slow progress is still a lack of modeling capability.

In December's Volcano FORCE conference, Volcano Engine President Tan to stay said that the core of the Agent landing or technology to be good; Wisdom Spectrum 2024 is also continuing to recruit technical talent, one of the purposes is to make Agent better by improving the modeling capabilities.

Throughout 2024, the biggest variable in big model technology is the law of the road to AGI - the Scaling Law is failing.

A paper published by OpenAI in 2020 states that model performance increases with the number of model parameters, the size of the training data, and the training computational resourceshow linear growth, which was the central point of the first generation of Scaling Law.

However, the actual results of a model with such parametric quantities as Llama-3.1-405B, after exhausting almost all publicly available datasets for training, are not satisfactory - the results of many review scenarios show that it is not very different from its own 70B model; and OpenAI has not released a general-purpose model in 2024, after GPT-4 base big model with significantly improved capabilities. Its GPT-5, which was exposed by the Wall Street Journal as far less effective than expected due to insufficient high-quality data and other reasons.

"GPT-5 hasCompletion of at least 2 rounds of training.It was months long each time, but new problems were encountered after each training session." The Wall Street Journal quoted people familiar with the matter as saying.

(Reported by the Wall Street Journal)

The pre-training recipe for scale failure, the solution given by OpenAI and Anthropic, is to boost model inference through reinforcement learning (RL).

In mid-September 2024, OpenAI went live with its first inference model-o1. At the end of December, an advanced version of the inference model, o3, went live. It demonstrated the ability of large models to solve complex problems in science, coding, math, and other fields, drawing followers.

In November, two months after OpenAI's o1 was released, there were at least nearly 10 teams in China reproducing o1.

But models of reasoning like o1 and o3 are also controversial.

"Measured a circle of domestic class o1 models, the effect is poor." An algorithm engineer engaged in model research told "Four Woods Relativity", "Ali's QwQ effect can be a little better. If o1 is 100 points, QwQ can reach 60-70 points."

Another voice questions the "narrow application scenarios" of o1 and o3. Outside of math and code, there are not many scenarios that it can be adapted to. If the cost is taken into account, the prospect of landing is not clear.

2025: Model Layer Convergence, AI Applications Take Off?

During this period of technological disorientation, the country's big modeling companies began to quietly diverge.

The price of training a model that performs close to Claude 3.5 Sonnet or GPT-4o is now not prohibitive.

DeepSeek's latest release, V3, cost $5.576 million to train.

However, it makes a point in its technical report to indicate that this price is only the cost of formal training, and does not include the cost of previous research and experimentation with algorithms, architectures, and data.

To really break through to the next generation of models, the big modeling companies have to pay a higher "price of trial and error".

OpenAI's experience side-steps this.

Back in 2022, before it hit the big time, it cost $540 million a year. And by 2024, The Information reports that OpenAI is feared to cost as much as $8.5 billion. Of that, training and reasoning costs account for $7 billion, and staffing costs $1.5 billion.

Screening the two elements of talent and arithmetic, I'm afraid that only four or five domestic companies will be left in 2025 that insist on training the underlying big model.

According to tech consultancy Omdia, the top domestic GPU buyers in 2024 will be Byte and Tencent. We've learned that Byte has been stocking up on cards internally "at all costs" since 2023, and it's still grabbing GPUs and GPU-enabled machines in 2024.

(Tech consultancy Omdia estimates that the domestic companies buying the most cards in 2024 will be Byte and Tencent)

The density of top talent is another weight that enhances the effectiveness of the model.

A head of the big model company executives have commented to us, byte beanbag big model of the effect of the high price to dig the former Tongyi thousand asked the technical head of Zhou Chang, only after the obvious upturn.

In line with Zhang Yiming's action, recently, another big man in the tech circle, Lei Jun, also began to personally go down to poach people.At the end of December, DeepSeek member Luo Fuli was revealed to be on the verge of joining Xiaomi's big model team.

"Shiki Relativity" also learned that since Lei Jun's Shunwei Capital invested in companies such as Smart Spectrum, Dark Side of the Moon, Baichuan Intelligence and MiniMax. Lei Jun also personally contacted the executives of six small tigers during the recruitment period.

"But the Offer offered by Xiaomi is not necessarily attractive." An AI investor said that although Xiaomi is trying to dig people, but counting the computing power resources and brand awareness, it is still difficult to compare with the "big grabbers" byte.

Among the Six Tigers, the companies thought to be making a firm push into the modeling layer are Smart Spectrum and Steps. They each announced large funding rounds of hundreds of millions of dollars in December.

There's also a surprising face appearing in the capital markets. Multiple sources say DeepSeek was moving forward with independent financing months ago.

(DeepSeek releases general-purpose model DeepSeek-V3 on Dec. 26, image source DeepSeek official website)

Grabbing people, grabbing cards, and "grabbing money," the players in the modeling tier are stocking up on food and drink for 2025.

After a fast and furious year, the early arbitrage space for the underlying big models came to an end.

For the remaining startups, the probability is that they will focus on AI applications.2024 At the end of the year, more than a dozen early-stage investment organizations told us that the focus will be on AI applications in 2025.

This trend is beginning to take hold.

The fastest projects to take large amounts of financing in the near future are almost all AI applications whose founding teams are executives of large companies or six tigers co-founding them. This type of project team background is shiny, the financing cycle is very short, usually by three or four well-known dollar funds together to grab the finalization of the first round of financing, the valuation quickly soared to the level of hundreds of millions of dollars.

The most recent example of this is the consumer robotics startup of Renjie Guo, the former CEO of Chasing China. We also learned that some high P's from big manufacturers and startups have recently been preparing for their startups and are making initial contacts with investors.

As for small and beautiful AI application teams, they won't be able to capture huge amounts of funding, but have a higher probability of reaping small investment checks of around $500,000 dollars. These teams typically have single-digit employees. They will make delicate functions/applications such as AI content creation, AI health monitoring, AI mini-games, etc. based on the big model. $500,000 to $1.5 million investment is enough to support them to run the first or second version of the product and get users and a part of the revenue.

These signs suggest that by 2025, the funding environment for AI app entrepreneurs may be much more relaxed.

But another problem is that whether it is AI Coding and consumer robotics projects by star entrepreneurs or AI creation apps and health monitoring by small teams of developers, they all seem to be a bit far from what they imagined Killer APP to be.

"Team background matters, data matters, growth matters." More than one AI investor summarized his investment criteria for AI applications this way.

This criterion, which mirrors that of the mobile internet era, seems to indicate that thinking about the Killer App has been put on hold.

Will 2025 tell us the answer?

This article was written by Shiki Relativity , source:Tencent Technology (developers of the QQ instant messaging platform)The original title: "Big Model "Six Tigers" 2024: Big Factory Strong, Hard to Find a Way".