Mushishi Sovereign AI Stack shipped

Self-hosted multimodal LLM infrastructure where 'client data never touches a cloud API' is enforced by routing, not policy.

github.com/MushiSenpai/mushishi-sovereign-ai-stack ↗  ·  vLLM · Nemotron NVFP4 · LiteLLM · llama.cpp · Tailscale · restic


One RTX 5090, 32GB of VRAM, and three fallback chains with three different guarantees — including a client profile that refuses to fall back, because for paid work, failing loudly beats degrading silently.

Measured: 275 tok/s on Nemotron NVFP4 via vLLM, 180K context at FP8 KV cache. The repo carries the full Decision Log, including the six hours of TensorRT-LLM debugging that ended in choosing vLLM — the dead ends are documented because they were the expensive part.

Posts about this project