wtf is kimi-k3?

Kimi K3 is a large AI language model released by Moonshot AI, with its full set of trained weights made publicly available rather than kept private behind an API only. It is described as the company's most capable model so far, and the README calls it an open, native multimodal, agentic model, meaning it can understand text, images, and video in one model, and it is designed to carry out multi step tasks with a degree of independence, not just answer single questions. Under the hood it is enormous: 2.8 trillion total parameters, though only about 104 billion of those are active for any given piece of input, using a mixture of experts design where the model picks a small subset of specialized subnetworks to use per token. It supports a context window of about 1 million tokens, meaning it can take in a very large amount of text, code, or conversation history at once. The README highlights three main strengths: long running coding work such as navigating large codebases and using terminal tools with minimal human help, agentic knowledge work like producing research reports with charts and interactive elements, and native understanding of images and video alongside text. The model weights are released under something called the Kimi K3 License, described as open for research, deployment, and further development, though this is a custom license rather than a standard one like MIT or Apache. The README includes a large technical table of architecture details for researchers, such as its number of layers, attention heads, and expert count, and points to a full technical report and blog post for deeper explanation. This is a research grade AI model release at its core, not a ready made application, so using it directly requires the infrastructure to run very large models. The full README is longer than what was shown.