The Chinese AI model and the real questions to ask about the new Kimi K3

Milan. In the rapidly evolving field of AI, rankings are updated on a weekly basis. However, when a Chinese model reaches the top of the performance rankings for LLM models (i.e. artificial intelligence systems capable of generating text, such as ChatGPT or Claude), the news is always perceived as a challenge to Western (or rather, American) technological leadership. Kimi K3, the Chinese open-weight model – that is, one whose parameters are made public and freely downloadable – with 2.8 trillion parameters released by Moonshot AI, appears to outperform even Anthropic’s much-celebrated Fable or the latest, ‘steroid-fuelled’ version of ChatGPT. The news comes as a surprise only to those who had not been following the developments surrounding Deepseek (which led to significant stock market corrections for Nvidia and the entire sector); consequently, it is the reactions coming from Silicon Valley and Washington that are proving most interesting.

Dean Ball, who recently moved from the Trump administration to head up strategic planning at OpenAI, has devoted a lengthy train of thought to the phenomenon, and between the lines one can sense a concern that relates less to Chinese technology itself than to the economic model that this technology is quietly building. Ball’s central argument is that open-weights models are inherently decelerationist, as they would discourage the investment from large companies needed to build the next generation of systems. Accelerationists who celebrate the opening up of these models’ weights (a sort of ‘open source’ but without the sharing of training data), he argues, are in fact doing so not so much for the speed of progress as for the ungovernability that openness produces – a refuge from regulation rather than a catalyst for discovery.

This is where the argument starts to fall apart, and it is no coincidence that rather sharp replies have appeared beneath his own post, turning the perspective on its head. If one accepts that scientific progress depends on the widespread dissemination of the most capable tools, then making, as is the fashion in Beijing, an advanced “cutting-edge” model available to anyone – an AI as a common good, to use Xi Jinping’s words – from university laboratories to start-ups without the budget for expensive licences, may well be the most efficient way to truly push the frontiers of knowledge, whilst a closed, Silicon Valley-style system may indeed attract capital to the laboratories developing it, but without guaranteeing that it will be applied on a large scale. Then there is an argument that affects Ball more directly: those who keep the models under lock and key become the arbiters of who can access the technology, replacing the institutions that currently fund science – universities and grant-awarding bodies – with a discretionary power that laboratories have never exercised before. The very same security measures with which Anthropic restricted access to Fable for certain biology researchers were cited – even in the replies to Ball’s tweet – as evidence that the closed approach also has its own limitations. The point, in short, is that ‘decelerationism’ – assuming it is a concept with exclusively negative connotations – is a term that depends entirely on what is meant by ‘acceleration’: if it refers to the concentration of capital in the hands of the few laboratories that can afford to train models costing billions of dollars, then the Chinese-style openness of models does indeed slow it down; if it is the speed at which knowledge spreads throughout society, openness could, on the other hand, be its most powerful driving force.

Ball’s underlying concern remains that open-source models may end up legitimising the Chinese vision of artificial intelligence as digital public infrastructure – provided by the state rather than sold on the market – a scenario he describes as ‘openly dystopian’. Here, the figures help to explain why Beijing has no qualms about pursuing this path. According to Stanford’s AI Index 2026, China holds 74.2 per cent of global AI patents, leads in scientific publications and citations, and in 2024 accounted for 54 per cent of global industrial robot installations, with the humanoid robotics sector growing at the same rate as consumer electronics. The United States, however, retains a profound structural advantage: private investment in AI reached $285.9 billion in 2025 – twenty-three times that of China – and American patents generate over half of all subsequent citations worldwide, a sign of quality that China’s sheer volume alone cannot compensate for. The performance gap between the two countries’ top models, according to the same Stanford report, has narrowed to 2.7 per cent. It is within this increasingly narrow margin – between Chinese volume and American influence – that the real contest is being played out, far more so than in the single, fleeting ranking won by Kimi.