modelship-cpu
ghcr.io/modelship-ai/modelship:latest-cpu
https://ghcr.io/modelship-ai/modelship
bridge
sh
false
https://github.com/modelship-ai/modelship/issues
https://github.com/modelship-ai/modelship
Self-hosted, OpenAI-compatible inference for the agentic era. Modelship runs your whole AI stack -- reasoning LLMs with universal tool calling, embeddings, speech-to-text, text-to-speech, and image generation -- as multiple models sharing your CPU behind a single OpenAI-compatible gateway. Built on Ray Serve with pluggable backends (llama.cpp, vLLM, Diffusers, plugins).
This is the CPU-only variant -- no GPU or NVIDIA Container Toolkit needed, works on any box (amd64 or arm64, including Apple Silicon hosts). For GPU support instead, use the separate "modelship-cuda" template.
IMPORTANT: create your models.yaml at the path below BEFORE starting the container -- it will not start without one. See https://github.com/modelship-ai/modelship/blob/main/docs/model-configuration.md and https://github.com/modelship-ai/modelship/tree/main/config/examples for the format and ready-made examples.
IMPORTANT: the default --shm-size=2g in Extra Parameters is sized only for a single small quick-start model. Ray sizes its object store to roughly 30% of the RAM it sees as available AT STARTUP -- if --shm-size is smaller than that, Ray crashes immediately. Raise --shm-size (roughly 30% of the RAM you want to give modelship) for bigger or multi-model stacks.
AI:
http://[IP]:[PORT:8000]/
https://raw.githubusercontent.com/modelship-ai/modelship-unraid/main/icon.png
--shm-size=2g -e MSHIP_UID=99 -e MSHIP_GID=100
CPU-only -- runs on any Unraid box, no GPU or NVIDIA Container Toolkit needed.
8000
8079
/mnt/user/appdata/modelship/models.yaml
/mnt/user/appdata/modelship/cache
INFO
true
modelship