Moonshot AI releases open-weights Kimi K3 amid distillation debate

In short: Moonshot AI has released Kimi K3, a very large open-weights AI model from China, and commentators are debating whether it was trained by copying ideas or outputs from US models.

What happened

Moonshot AI introduced Kimi K3 through an API in mid-July 2026 and then released its “open weights” under a Modified MIT-style license. Open weights means other people can download the model’s core files and run it themselves, like getting the recipe instead of only being able to eat at the restaurant.

Kimi K3 is a Mixture-of-Experts model, which is a design where the system has many specialist parts but only uses a small set at a time (like a company with hundreds of employees, but only the right 16 join each meeting). Reports put it at about 2.7 to 2.8 trillion parameters, a rough measure of model size.

In public benchmark results, K3 does especially well on coding and “agentic” tests, which measure whether an AI can take steps to complete a task using tools and instructions. On the Artificial Analysis Intelligence Index, it reportedly scores 57, placing around third or fourth globally, behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol.

The New York Times podcast Hard Fork discussed a key question: was K3 built by “distilling” American models. Distillation usually means training one model to imitate another model’s answers. Based on public reports so far, there is no hard evidence that K3 directly copied any specific US model’s internal weights or code, even though many labs do train on broad internet data where model-generated text can appear.

Why it matters

Open-weights releases can spread quickly because more companies and researchers can build on them. The debate also highlights a bigger issue: it is often unclear what data and other models were used to train today’s top systems.

Source: NYTimes