How to build custom Docker image for AI models

What this does

Packages a customized AI model and its runtime into a self-contained Docker image that runs anywhere Docker is available. After this guide the image is built, tagged, and confirmed to load the model successfully.

Steps

1. Create the Dockerfile. The Dockerfile defines the base runtime, model injection point, and startup behavior.

   FROM ollama/ollama:latest
   COPY ./models/llama3-8b-instruct-q4_K_M.gguf /root/.ollama/models/
   CMD ["serve"]
   

   Place this file alongside a models/ directory containing the GGUF file.

2. Build the image with an appropriate tag.

   docker build -t ollama-custom:llama3-8b-q4 .
   

   Expected output: Docker build finishes with Successfully built <image-id> and Successfully tagged ollama-custom:llama3-8b-q4.

3. Run and validate the container.

   docker run -d --name ollama-custom -p 11434:11434 ollama-custom:llama3-8b-q4
   sleep 15 && curl -s http://localhost:11434/api/tags
   

   Expected output: JSON listing model names available in the container.

4. Push to a registry if needed. For multi-host deployments.

   docker tag ollama-custom:llama3-8b-q4 myregistry.example.com/ollama-custom:llama3-8b-q4
   docker push myregistry.example.com/ollama-custom:llama3-8b-q4
   

   Expected output: Layer upload progress followed by Pushed confirmation.

Verification

docker run --rm ollama-custom:llama3-8b-q4 ollama list
# Expected: Table row listing the injected model with size and quantization

Common failures

• COPY failed: file not found — The model file is not in models/ relative to the build context. Verify with ls models/ from the build directory.
• connection refused when querying the API — The container failed to start. Run docker logs <container-id> to inspect the startup error.
• exec format error — Architecture mismatch. Verify with docker inspect <image> | grep Architecture.
• Model not listed after build — Ollama does not auto-import copied GGUF files. Use OLLAMA_MODELS env var or run ollama create inside the container.
• Image too large — The model file bloats the Docker image. Use a .dockerignore to exclude source files and multi-stage builds to minimize layers.

Related guides

• How to mount GPU in Docker container
• How to persist Ollama data in Docker
• Course Ollama Deep Dive