Kimi K3 represents Moonshot AI's push into the 3-trillion-parameter open-source frontier, combining 2.8 trillion total parameters with a Kimi Delta Attention and Attention Residuals hybrid architecture. This design pairs a linear attention mechanism with residual attention pathways, then layers on a highly sparse Stable LatentMoE framework that activates 16 of 896 experts. The result is a model that scales MoE sparsity aggressively while preserving the ability to attend over very long sequences, all wrapped in a native multimodal foundation that handles visual input directly rather than treating vision as an add-on.

Positioned for long-horizon coding, complex knowledge work, and deep reasoning, Kimi K3 is described by its creators as the first open-source model to reach the 3-trillion-parameter class. The 1M-token context window and native visual understanding make it well suited to projects that need to reason across large codebases, lengthy documents, or mixed image-and-text inputs. Practitioners looking for a self-hostable flagship with frontier-scale capacity, extended context, and visual reasoning will find Kimi K3 aimed squarely at those workloads, with additional architecture, training, and benchmark detail promised in the accompanying technical report.