Your AI advantage is evaporating

Chinese AI companies are throwing it away. After the launch of the promising Kimi K3, now Alibaba has just introduce Qwen 3.8a gigantic model that according to those responsible is so powerful that only Fable 5 manages to surpass it. If that is confirmed, we will be facing a potential turning point in the AI ​​race. One in which China will no longer be months away.

A fearsome rival for GPT-5.6 and Fable 5. Qwen 3.8 Mac has 2.4 trillion parameters, placing it only “slightly” behind the also gigantic Kimi K3 which we learned about last week and which has 2.8 billion parameters. Developers can now begin testing the model through Alibaba’s Qoder programming platform, and those responsible for it have announced that they will release their weights soon. Alibaba’s message is quite a challenge, because they claim that only Fable 5 manages to exceed their capacity. That’s saying a lot… or maybe not.

All unknowns. The truth is that this is a preliminary presentation because the availability of the model is very limited. In fact, the company has not even published internal benchmarks and that makes it difficult to validate such claims. Some users have shared theoretical results of the behavior of Qwen 3.8 and according to said data it does seem that we are facing a prodigious model, but one that is also unknown details such as cost or efficiency.

Meanwhile, Kimi K3 already shows his power. The launch comes just days after Kimi K3 have managed to overcome to Fable 5 and GPT-5.6 Sol in certain specific tests of the Arena.ai benchmark platform. It also surpassed Opus 4.8 in the text generation test classification, something equally striking, especially if we take into account that the Chinese model costs 40% less. The demand for Kimi K3 has been so high that Moonshot AI has had to announce that for now need to cancel new subscriptions to deal with the enormous amount of requests that their servers are receiving.

GLM 5.2 had already struck its own blow. We have had a very busy summer in terms of new AI models from the Far East. In June The big surprise was carried out by GLM-5.2the open Zhipu AI model that with its 744 billion parameters managed to be on par with Opus 4.8 at a fraction of its price. This model left behind all its Chinese competitors at the time (MiniMax, DeepSeek V4 Pro, Kimi K2.6), and began to make it clear that China had “a Mythos at home”.

The pattern that begins to be drawn. Neither GLM-5.2, nor Kimi K3 later nor now Qwen 3.8 need to be the best models in the world to change the rules of the game: for companies and developers, having access to a model that performs almost like the best, that costs (sometimes much) less and that can even be installed in their own systems because they are open models can be more attractive than paying for the very expensive frontier models from OpenAI or Anthropic. Each launch actually calls into question the exorbitant amounts these companies are investing in ever larger data centers.

Alibaba scores the same. The Chinese company will have to confirm with data that Qwen 3.8 really is as good as they say, but for now they have already achieved a relevant victory: the Beijing government approved Apple Intelligence, and will allow iPhones, iPads and Macs to use said technology in the Asian giant with a very special detail: the AI ​​models used will be from Alibabanot from Google.

I distill, you distill, he distills. And while all this is happening, Anthropic accuse Moonshot and other Chinese companies to “distill” their models on an industrial scale to develop their own proposals. That is to say: they claim that if the Chinese models are so good it is because they are doing a kind of reverse engineering on models like Claude Opus 4.8, for example. But the fact that many Chinese models are open has also allowed US companies to take advantage of these models for all types of projects, including specific models. like Composer 2.5 from Cursor or the new Inkling from Thinking Machines that It is based on DeepSeek-V3.

The gap disappears. The launch of these models demonstrates that the gap between Chinese and American AI models has narrowed significantly. Not long ago it was said that it took China about six months to match the capacity of US models, but these developments make it clear that now it only takes a few weeks before models comparable in performance appear, but with open weights and a reduced cost. If the race was already getting interesting, now it gets fascinating.

In Xataka | A company decided to give the keys to its systems to an AI agent. The AI ​​agent deleted its entire database