1 trillion parameters on 8 GPUs. Here's what that looks like. We deployed Kimi K2.6 on Hyperstack - @Kimi_Moonshot's open-weight agentic model. In this video: → vLLM serving on 8x NVIDIA H100-80G PCIe → 595 GB of INT4 weights loaded from ephemeral NVMe in ~6 minutes → Autonomous multi-step refactoring with self-correction → Coding-driven design - single prompt to working website → Local backend for Claude Code, OpenClaw and Kimi Code CLI 32B active parameters per token. 256K context window. 300 sub-agents in a single run. Full tutorial on our blog: bit.ly/4cFJVLF #KimiK2 #MoonshotAI