Deploy DeepSeek‑R1 on AutoDL with Ollama: A Step‑by‑Step Guide

DeepSeek‑R1, the latest open‑source AI large model that claims performance comparable to OpenAI o1, can be privately deployed to retain data privacy, optimize compute resources, gain flexibility, customize algorithms, control costs, and simplify maintenance.

For private deployment, the author selects an AutoDL GPU cloud instance (e.g., a 4090 GPU with 24 GB VRAM, 16 CPU cores) because local machines may lack sufficient performance; AutoDL offers hourly pricing (≈0.6‑3 CNY) and convenient environment provisioning.

After registering on AutoDL, the server is accessed via SSH (using FinalShell) where a Docker‑based container runs. System utilities are installed with commands such as sudo apt‑get update, sudo apt‑get install lshw net-tools systemd systemctl.

Ollama, a free open‑source tool that bundles pre‑installed LLMs, is installed by executing curl -fsSL https://ollama.com/install.sh | sh. The Ollama service is started with systemctl start ollama.service and its status checked via systemctl status ollama.service.

To expose the service externally, the author modifies /etc/systemd/system/ollama.service adding Environment="OLLAMA_HOST=0.0.0.0:6006", reloads the daemon ( sudo systemctl daemon-reload) and restarts the service. Port 6006 is then verified with netstat -an | grep 6006.

The DeepSeek‑R1 14B model is pulled using the command OLLAMA_HOST=127.0.0.1:6006 ollama run deepseek-r1:14b. The pull completes quickly thanks to AutoDL’s network acceleration, and the model responds promptly on the 24 GB VRAM server.

For a user‑friendly UI, OpenWebUI is run in Docker:

docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui --restart always ghcr.io/open-webui/open-webui:main

. The OpenWebUI settings are updated to point to the Ollama endpoint http://localhost:6006/v1/chat/completions, enabling interactive conversations with DeepSeek‑R1.

Finally, the author reminds users to shut down the AutoDL instance when not in use to avoid unnecessary charges.