Uno GPU wants private models to sleep between thoughts

Field notes

Uno GPU describes a snapshot-based inference workflow: a model state sits beside the GPU, streams into VRAM on request, serves through an OpenAI-compatible endpoint, and returns to sleep when idle. Its worked example uses Qwen3 27B, a 27 GB state, and an H200 slice with 35 GB of VRAM; the page states a 0.58-second resume and 0.60-second first-token path. The GPU offering is presented as early access, while the underlying snapshot mechanism is attributed to Uno Cloud production use.