Moonshot AI has released the Kimi K3 model, a large language model with 2.8 trillion parameters. The company also open-sourced the infrastructure behind the model, a move that could reshape how developers and researchers build on top of such systems.
What the Kimi K3 model brings
The Kimi K3 model’s 2.8 trillion parameters place it among the largest publicly known language models. While many frontier models remain closed or partially open, Moonshot AI has chosen to release the underlying infrastructure — the code, training tools, and deployment framework — alongside the model itself. That means other teams can inspect, modify, or adapt the system without starting from scratch.
Why open-sourcing the infrastructure matters
Open-sourcing the infrastructure, rather than just the model weights, gives developers access to the full pipeline used to train and run Kimi K3. This includes the data processing, distributed training setup, and inference optimizations. For researchers and smaller AI labs, that level of transparency can lower the barrier to experimenting with large-scale models. It also allows the community to verify claims about the model’s performance and safety.
The company did not specify a license for the open-sourced components or detail any restrictions on commercial use. Those details will likely determine how quickly the infrastructure gets adopted.
What comes next
Moonshot AI has not announced a release date for the model weights or a public API. The open-sourced infrastructure is available now, but the model itself may follow later. Developers and researchers will be watching for the exact terms of use and whether the company plans to offer hosted versions of Kimi K3.
