Imp

About

Imp is an LLM inference server built only for NVIDIA Blackwell GPUs (GeForce RTX 50 and RTX PRO). It exposes OpenAI and Anthropic-compatible APIs with tool calls and runs from a single Docker container with no local CUDA toolkit or Python needed. The author reports 96.3 tokens per second for Qwen3.8-27B on one RTX 5090, ahead of vLLM, and can serve a model larger than the card's memory by keeping experts in host RAM.