Kimi K3: The counterattack of AI in China from closed to open source

Subtitle: Benchmarking the Silicon Valley paradigm and exploring how tools reshape developers

On the surface, Kimi K3's open source is a battle for model performance, but at the bottom, it is actually a reverse penetration of paradigm: China's Internet community has begun to actively move closer to Silicon Valley's "open source is ecology." This change far from just giving developers one more model to play with, but also bringing China's AI ecosystem into the home of global collaboration and handing over the right to define tools. This is also a counter-intuitive phenomenon I noticed-China is good at "fast + closed", but chose "slow + open" in the AI round, and its attitude is quite high.

After taking a look, you will understand that this is not the Kimi family who wants to "show off their muscles," but the underlying paradigm of the entire China Internet is being manipulated. In the past ten years, domestic large models have been built behind closed doors, emphasizing their own data, their own platforms, and their own users. Now that the source is suddenly released and even the training code is spilled out, there is a new understanding of the "cost of efficiency". To put it bluntly, this is not a technological upgrade, but an ecological upgrade.

QKPFX0 Reverse penetration of QK open source paradigm

Look at the facts first. The open source protocol offered by Kimi K3 is Apache 2.0, which is commercially available with no additional conditions. Directly benchmark OpenAI's GPT-3.5, Llama-3, Mistral and other mainstream Silicon Valley models. The parameter quantity is 26B, inference speed, context length, Huggingface compatibility, and even the "full training process" are given. This full-link open source is not only "you can use it", but also "you can change it."

In the past two years, the open source actions of China's major models have been "symbolic" open source: the models are thrown out, the code is hidden, the training process is missing, and no one cares about issues raised by the community. This time, Kimi K3 even threw out data sets and training scripts, and even encouraged community forking and secondary development. Don't play half-open and half-closed routines.

What's interesting about this is that the paradigm of open source models in Silicon Valley has long been formed-mainstream models are all hung on Huggingface and GitHub, developers pull requests from each other, and the ecosystem is driven by the community. China has always been "platform-driven" in the past: you have to use my API and my platform to enjoy modeling capabilities. Now Kimi K3 directly moves the paradigm and no longer does "platform binding".

What does this represent? In essence, China's AI tools voluntarily abandon "monopolistic efficiency" and turn to "collaborative innovation." The cost of efficiency has begun to emerge in the AI era: closed-loop products are made quickly, but ecology is not done much.

The above figure is not a statistical statistic, but a subjective estimate of paradigm maturity. Kimi K3's "95%" means that it has brought ecological openness to the mainstream level of Silicon Valley-not only are models available, communities can be changed, and training processes have no thresholds.

How ## tools reshape developers

It is no exaggeration to say that tools determine the boundaries of developers. The traditional way of playing the Internet in China is "tool is platform", and developers are either API callers or second-level developers. Few people can directly participate in the optimization of the underlying model. Kimi K3 threw out training code, data sets, and reasoning processes this time, which means that developers can no longer just "call" but "define".

To put it bluntly, the openness of tools determines the innovation space of developers. In the past, the collaborative ecosystem brought about by Llama-2 and Mistral on Huggingface was where hundreds of thousands of developers and thousands of companies simultaneously did fine-tuning and vertical adaptation. In the previous model in China, you could only use its API and couldn't change the bottom layer. Putting your own needs on it would be considered "innovation." Now Kimi K3 directly follows the "Silicon Valley paradigm", allowing developers to train, tune, and deploy themselves and are no longer limited by the platform.

What does this mean for China developers? Essentially, it is from "passive consumption" to "active creation". In the past, everyone could only wait for big factories to upgrade their models, and write some tips themselves as "innovation". Now you can fork Kimi K3, add new functions yourself, and even let big manufacturers absorb your results. Tools become ecosystems, not platforms.

The reason why this is important is that China developers have always been kidnapped by "platform efficiency": you either make plug-ins or do API integration, and there are very few opportunities to really touch the bottom line. Now that Kimi K3 has let go of its source and left the power of "definition tools" to the community, developers have the opportunity to stand in the leading role of the ecology for the first time.

The paradigm difference between China Internet and Silicon Valley

The advantages of China's Internet have always been "fast + closed": products are launched quickly, users are growing rapidly, and closed-loop control is strong. But the way the AI ecosystem plays has changed, and Silicon Valley has won the ecosystem by relying on "slow + open". Models such as Llama-3, Mistral, Gemini, and Stable Diffusion all rely on the open source community to make countless derivative products, and their ecosystems are far richer than any single platform. China's major factories used to implement closed-loop operations, resulting in low community activity and slow ecological growth.

The essence behind this matter is "how tools reshape people who use tools." Silicon Valley's open source paradigm has turned developers into the protagonists of the ecology; China's closed paradigm has turned developers into "platform outsourcing teams." Kimi K3 open source is equivalent to pulling developers from the "outsourcing team" back to the "ecological core". This is the real counterattack.

To give a cross-border analogy: In the early days of Android and iOS, platforms both wanted to hold the ecology firmly in their hands. Later, Android opened its source code, and developers from all over the world poured in, making the ecosystem far richer than closed-loop iOS. China's Internet has always followed the iOS route, and as a result, the AI ecosystem has been trapped by "efficiencyism." Kimi K3 learned Android this time and went to the open source home, which was equivalent to handing ecological leverage to the community.

steelman: Opposition Views and Responses

Of course, some people will say that open source is just a "show." China's big model lacks real data and algorithm breakthroughs, and no one will use it even if it is open source. The opposing views are usually: "If you let go of the source, the community will have no motivation, and the ecology will still be stagnant; China developers are accustomed to using platform services and are unwilling to mess with the underlying code; big factories put open source only for marketing, and real innovation must rely on closed loop."

This statement actually makes some sense. Engineers in China are indeed more accustomed to "easy-to-use APIs." They have a high threshold for tinkering with underlying code, and the community atmosphere is not as strong as Silicon Valley. In the past, Dachang's open source projects, the issue area was abandoned all year round, the PR was ignored, and the ecological activity was indeed low.

But this time the Kimi K3 is different. First, the model itself is practical enough, not a "semi-finished product", can be directly deployed and commercially available. Second, the training process is complete, the dataset and code are all open, and developers can play with it themselves. Third, the community has become active: GitHub stars exceeded 4000 in the first week of its release, and people were following up on PR, issue, and discussion forums. The most important thing is that the technical atmosphere in the domestic AI circle is changing: more and more small and medium-sized teams and startups are beginning to fine-tune their models and no longer focus on the APIs of big manufacturers.

I bet this will work-as long as the tools are truly open, the ecology will eventually improve. Closed-loop efficiency can outperform for a while, and open ecology can win forever.

Ecological Reinvention

Last year, I met a small team working on AI projects. They have been using Dachang's cloud services for speech recognition. The API call cost is high and its functions are limited. If they want to upgrade, they can only wait for the Dachang to launch a new interface. After Kimi K3 was open source, these people directly pulled the code, trained the model themselves, and tuned it to their own scenarios. As a result, the performance increased by 20% and the cost was cut in half. Team members have changed from "API dispatchers" to "model engineers", and there is a lot more room for innovation.

This incident happens repeatedly in China's Internet circles. Only when developers change from "consumption tools" to "creation tools" does the ecological leverage point truly emerge. In the past, everyone waited for a big factory to upgrade, but now they can change the bottom layer by themselves. This is a typical scenario for a tool reshaping person.

The paradigm shift behind ## commercial accounts

In the final analysis, China's major manufacturers are not "feelings" to open source, but because commercial accounts cannot be calculated. The business model of the basic AI model is difficult to support by API revenue. OpenAI, Google, and Meta have all discovered this: the model itself relies on open source to make the ecosystem, and ultimately it is the ecological scale, not API orders. China's major factories used to rely on "closed-loop + platform" to collect money. As a result, the ecology did not do much and innovation was broken. Learning to open source in Silicon Valley now is equivalent to migrating the business model to "ecological harvesting."

But there are risks. After open source, the model will be forked, the ecology will be fragmented, and the dominance of large manufacturers will be weakened. China's Internet is used to "making its own decisions." Now it has to learn from Silicon Valley's "letting the community make the decisions." Whether it can really hold on depends on Kimi K3's community operations.

The real problem with ## is not model performance, it is paradigm shift

To put it counter-intuitively: Regarding Kimi K3 open source, mainstream discussions are all about "performance benchmarking GPT-3.5". In fact, the real problem is not performance, but paradigm shift. China's AI circle has always been entangled in "technological catch-up," but this time when open source is pursuing, it is not the parameter quantity, but the ecological paradigm.

Golden sentence time--"Tools determine ecology, and ecology determines the future." China's AI community has finally begun to learn to hand over the definition of tools to developers rather than firmly holding them in the hands of the platform. This is the key to counterattack.

Having said that, whether the open source ecosystem can hold up and grow large depends on whether developers are willing to take the initiative to create rather than wait for major manufacturers to upgrade. If China's Internet community can transform the tradition of "fast + closed" into "slow + open," Kimi K3 will become a real paradigm counterattack. If we still go back to the old path and turn open source into "symbolic", this wave will just pass by.

Thinking along this line of thinking, leave a question: ** When tools are truly open, will developers become smarter, or will they just roll on new platforms? **

This is interesting.