It seems like yesterday when DeepSeek R1 called into question an idea that many took for granted: that the race for advanced artificial intelligence It still had a clear owner in Silicon Valley. The emergence of the Chinese model helped trigger a massive sale of technology and led NVIDIA to suffer a loss daily capitalization unprecedented until then. As the months passed, that image lost intensity, but the message remained: the Chinese technological ecosystem was not willing to limit itself to keeping pace with the United States.
The following notice now has a different name: Kimi K3. Moonshot AI has just presented a model with 2.8 trillion total parameters that, as soon as it arrived, was placed at the top of Frontend Code Arenaahead of some of the most powerful proposals from Anthropic and OpenAI. But the story is not limited to a classification: developers and fans are already using it to create interfaces, games and recreations that anyone can see and, in some cases, try. That’s where this article really begins.
It is worth dwelling on the details of that classification. At the time of writing, Kimi K3 reaches 1,679 points in Frontend Code Arena, ahead of Claude Fable 5with 1,631, and GPT-5.6 Sol xHigh, with 1,618. The improvement compared to the previous generation is also striking: Kimi K2.6 was in 18th placewhile his successor leads six of the seven domains evaluated. For now, Arena maintains the label of preliminary result, so it is convenient to read this position as a very significant photograph, but still susceptible to change.
We are not facing a universal programming exam, but rather a very specific test. Frontend Code Arena compares web applications created by different models and lets users evaluate which one solves the task better, which one works more reliably, and which one presents a better experience. That approach is especially useful for measuring visible and practical capabilities, but it also has obvious limits. That Kimi K3 leads here tells us a lot about its frontend performance, although it doesn’t automatically allow us to extend that advantage to complex repositories, backend, mathematics, or general reasoning.
Outside of this specific terrain, photography remains favorable, although more balanced. Vals AI places Kimi K3 second among 38 models, with 74.70%just behind Claude Fable 5, which reaches 75.14%, and above GPT-5.6 Sol, with 73.12%. Artificial Analysis also places it among the most advanced systems in its classification, with 57 points and third place overall.
Where Kimi K3 seems to feel most comfortable is in tasks that combine programming, visual context and several chained steps. Arena supports its ability to build web interfaces, while Vals AI also records high performance in agent programming tests. Moonshot adds that the model can traverse large repositories, use terminal tools, and review screenshots of its own work to correct the output on the fly. That last capability, which the company calls “vision in the loop,” helps explain why it excels at transforming visual references into interactive products.
There are also several cautions before interpreting Kimi K3 as a definitive victory. Moonshot presents it as an open weight model, but those files have not been published yet and the company promises to release them no later than July 27. Nor should we confuse this openness with complete open source, because details about the license and the rest of the system are still missing. Its 2.8 billion total parameters belong to a sparse architecture that activates 16 of its 896 experts. The company itself recommends configurations with 64 accelerators or more, very far from what a conventional computer can offer.
The community reaction helps understand why Kimi K3 is attracting so much attention. One of the most striking examples is a recreation of macOS 27 which works within the browser and which its creator attributes to a swarm of model agents working for about three hours. They add to it Ballista, an interactive panel with a 3D balloon and several comparisons against Claude and GPT. They are not independent benchmarks, but demos shared by their own creators, but they allow you to see what kind of results the model is producing outside the tables.
To create something like the macOS simulation or the ballista game, we don’t need to model every element by hand from scratch. We can describe the resultattach a reference and commission Kimi to build a functional application, for example with HTML, JavaScript and various graphics libraries. The project is then tested, modified, and finally published or recorded for sharing. Kimi K3 can be used from Kimi.comKimi Work, Kimi Code or tools connected to its API, although it is not confirmed which specific environment was used in several of the examples we have seen.
It is still early to turn this launch into a definitive change of leadership. Fable 5 and GPT-5.6 Sun They are still ahead in several evaluations, the Kimi K3’s weights are not yet available and many of its capabilities will have to be verified with more time. Even so, what we have seen is already difficult to ignore: a Chinese company can compete for leading positions, offer competitive results and get the community to transform that capacity into real applications almost immediately. The race continues, but the margin between its main protagonists seems increasingly narrower.
Images | Kimi | Screenshot
In Xataka | China has a plan to win the AI war against the US. And DeepSeek is its champion
GIPHY App Key not set. Please check settings