Kimi K3 is an open-weight flagship model from Moonshot AI that pushes into multi-trillion parameter territory, with 2.8 trillion total parameters making it the first open release in the 3-trillion-parameter class. Its architecture is built on Kimi Delta Attention, a hybrid linear attention mechanism, paired with Attention Residuals, and scales MoE sparsity through a Stable LatentMoE framework that activates 16 of 896 experts. The model is designed for frontier intelligence scenarios spanning long-horizon coding, knowledge work, and reasoning, and pairs that scale with a one-the cataloged API limit and native visual understanding for image-grounded tasks.

Because full weights have been released on Hugging Face under Moonshot AI, Kimi K3 is well suited for self-hosted deployments, custom fine-tuning, and research workflows where openness and scale both matter. The combination of hybrid linear attention with residual attention pathways targets efficient long-context inference, while the visual encoder lets the same model handle documents, screenshots, and diagrams alongside text. In practical terms, it fits teams building agentic coding assistants, large-context knowledge systems, and multimodal reasoning tools that need a strong open foundation rather than a closed API endpoint.