Overview
- Moonshot published full Kimi K3 weights, a 47‑page technical report, and parts of its runtime stack on July 27, making the 2.8 trillion‑parameter Mixture‑of‑Experts model broadly inspectable and available for download.
- K3 is multimodal with a one‑million‑token context window and sparse MoE activation that Moonshot says uses about 104 billion active parameters during inference to cut per‑token cost.
- Practical self‑hosting remains hard because the weights total roughly 1.4 TB and Moonshot recommends large distributed GPU clusters with many high‑end accelerators, so most users will rely on third‑party hosts such as Telnyx which already added K3 to its inference API.
- Moonshot released the model under a custom Kimi K3 licence that permits wide research use but requires separate commercial agreements and branding or disclosure obligations for large ‘model‑as‑a‑service’ providers.
- U.S. officials and Anthropic have publicly alleged K3 used outputs from U.S. frontier models during development; Moonshot denies this and the dispute is driving renewed debate over export controls, IP rules, and limits on open‑weight releases.