Moonshot AI drops Kimi-K3, an open 2.8-trillion parameter model

Moonshot AI just released Kimi-K3, making it the first open-weights 3T-class AI model available to developers.

Moonshot AI dropped the full model weights for Kimi-K3 on GitHub, pulling in over 4,600 stars this week. The model features 2.8 trillion total parameters built on a Mixture-of-Experts architecture, though it activates just 104 billion parameters per token. It handles text, images, and video natively, packs a 1-million-token context window, and claims a 2.5× scaling efficiency bump over Kimi K2.

Why it matters: Frontier-grade open models are rare, especially at this size. Developers get access to an open model built for long-horizon engineering—capable of handling complex repo navigation, terminal tool orchestration, and hardware tasks like chip design without relying on closed APIs.

Here's the gist on how it runs: across its 93 layers, a sparse gating system selects only 16 out of 896 experts per token. You can find the weights available now under the custom Kimi K3 License.

Open source isn't just chasing closed flagships anymore; it's matching their scale line for line.