Moonshot AI made the complete weights of Kimi K3 downloadable today, and the headline number is genuinely unusual: 2.8 trillion parameters in a mixture-of-experts architecture, which makes it the largest open-weight model anyone has shipped to date, alongside a one-million-token context window that accepts text, image, and video input. What makes this more than a benchmark flex is that it's a full weights release rather than an API-only announcement: Moonshot published the model on Hugging Face gated behind a license click-through, with MXFP4-quantized weights bringing the download to roughly 594 gigabytes, which is large but genuinely tractable for a team with the hardware to run inference on a model this size. Reports differ slightly on the exact license terms, some describing it as a modified MIT license and others as Apache 2.0, so anyone planning to build a commercial product on top of it should read the actual license file rather than trust secondhand summaries. The practical significance for developers is straightforward: teams that need a frontier-capability model but can't or don't want to depend on a proprietary US vendor's API, whether for data sovereignty reasons, cost control at scale, or simply wanting to fine-tune and modify the weights directly, now have a genuinely capable option to self-host. That tradeoff comes with real infrastructure cost, since running a 2.8-trillion-parameter MoE model requires substantial GPU memory and orchestration even with quantization, so this isn't a laptop-friendly release. But for organizations already running serious self-hosted inference infrastructure, it meaningfully raises the ceiling on what's available outside the closed-model ecosystem, and it continues a trend where Chinese labs are shipping open-weight models that increasingly compete with proprietary frontier systems rather than trailing a generation behind them.