Moonshot AI releases Kimi K3 and backs open AI model access

In short: Moonshot AI released Kimi K3 and used the launch to argue that open AI models are more transparent and accessible than closed ones.

What happened

Moonshot AI, a Beijing-based startup, said on July 16, 2026 that it released Kimi K3, a 2.8 trillion-parameter AI model. Parameters are the many small settings inside a model that help it produce answers, like the number of adjustable knobs in a very large machine.

Coverage described Kimi K3 as the largest open-source or open-weight model by parameter count at the time. “Open-weight” means the company shares the model’s weights, which are the saved settings the model uses to produce results. It is closer to sharing the full recipe than only selling the finished meal.

Reports said the K3 release included the full weights, a technical report, and the training infrastructure, which is the software setup used to build and run the model. This is meant to let developers download it, run it on their own computers or servers, and modify it.

Why it matters

Moonshot and other commentators argue that open models can be safer because more people can inspect them, test them, and control how they are used locally. That is a transparency and access argument, not a settled scientific proof that open models are always safer. Open access can also make powerful systems easier for bad actors to misuse, even if the sources here focus more on the benefits.

The release also fits a broader pattern in China. Several Chinese firms, including DeepSeek, Z.ai, Alibaba, and Moonshot, have released open-weight models in 2025 and 2026, while many leading US labs keep their most advanced systems closed.

For more on the company, see Moonshot AI.

Source: NYTimes