Overview
Operational reality of downloading and serving frontier open-weight models on private clusters for data-sovereignty, vs using China- or US-hosted APIs.
Timeline
- 2026-07-27: Scheduled kimi-k3 full weights (~2.8T MoE, MXFP4); recommend 64+ accelerators (2026-07-26-kimi-k3-open-weights-july-27-release, 2026-07-26-kimi-k3-moonshot-official-blog)
- 2026-07-22: vllm day-0 serving prep for K3 (2026-07-26-vllm-kimi-k3-preview)
Key Players
- moonshot-ai
- vllm
- hugging-face
- nvidia / AMD as accelerator vendors
Analysis
Largest open-weight drops raise sovereignty upside and hardware floor simultaneously; fake pre-drop download links are a recurring risk.