Two Chinese companies released or previewed multi‑trillion‑parameter large language models within a single week. Alibaba previewed Qwen3.8‑Max‑Preview at the World AI Conference in Shanghai, presenting it as a 2.4 trillion‑parameter model that the company says is second only to Anthropic's Claude Fable 5. Three days earlier, Moonshot released Kimi K3, a model with 2.8 trillion parameters.
Both projects have emphasized that the model weights will be opened, in contrast to the approach taken by several U.S. labs that keep their largest models closed.
Capabilities reported
According to Alibaba, Qwen3.8‑Max‑Preview is the first Qwen model above a trillion parameters that can handle images, video and documents in addition to text. Moonshot's Kimi K3 likewise sits in the multi‑trillion parameter class. The announcements indicate both models are trending toward multimodal functionality and that developers plan to make weights available for download or external use.
Why this matters
Open weights change how powerful AI models can be controlled. If model weights are publicly available and reproducible, they cannot be easily embargoed, metered through a single provider, or restricted by export controls. That reduces the effectiveness of centralized barriers to distribution and makes near‑frontier capabilities more widely accessible.
Observers note that this does not necessarily overturn which country produces the single most capable model. The U.S. may still host the frontier system in terms of peak performance. However, by making very capable models open and widely distributable, the baseline level of available capability becomes global and harder to contain.
Implications and risks
The practical consequence is strategic: governments can attempt to keep the most advanced models behind access controls, but they are less able to prevent the spread of sufficiently capable alternatives once those models' weights have been released. Open‑weight models can be copied, run offline, and integrated into local systems, limiting the reach of export rules or provider‑side restrictions.
Conclusion
Alibaba's Qwen3.8‑Max‑Preview (2.4 trillion parameters) and Moonshot's Kimi K3 (2.8 trillion parameters), announced within days of one another, highlight a shift in tactics. Rather than competing purely to hold the single smartest model, some Chinese actors appear to be prioritizing broad, inexpensive, and hard‑to‑restrict distribution of near‑frontier AI by opening model weights, a development with meaningful regulatory and strategic consequences.
