Kimi K3.1 Pro: a 2.8 trillion-parameter open-weight model with 1M context

Moonshot's Kimi account (@Kimi_Moonshot) introduced Kimi K3.1 Pro on Friday. The post describes a 3T-class open-weight mixture-of-experts model with 2.8T total and 104B active parameters, a 1M-token context window and multimodal input. It also says the model is aimed at full-project coding for Unreal, Unity and Godot. The post drew about 17,700 likes.

In a mixture-of-experts model only a slice of the parameters runs for each token, which is why the active count (104B) is far smaller than the total. Open-weight means the model files are meant to be downloadable, though the post does not cover licence terms or benchmarks.

Why it mattersVery large open-weight models with long context are the main competition for closed coding models, and game engines are a concrete target. Wait for independent tests before taking the capability claims at face value.