How to Run DeepSeek-R1 on a Linux Cloud VPS with Ollama (Zero GPU Needed)

The common assumption is that self-hosting cutting-edge reasoning LLMs like DeepSeek-R1 requires renting an \$800/month NVIDIA H100 GPU instance. In 2026, that assumption is completely obsolete. With 4-bit and 8-bit GGUF quantization and modern CPU AVX-512 vector execution, you can run DeepSeek-R1 on a standard high-RAM KVM Cloud VPS at 35 to 60+ tokens per second — for a fraction of the cost.

In this walkthrough, we will configure an Ubuntu KVM VPS, install Ollama with systemd, deploy DeepSeek-R1 distilled reasoning models, set up an OpenAI-compatible API endpoint, and secure a multi-user Open WebUI dashboard.

Why Run DeepSeek-R1 on a Private Cloud VPS?

When you rely on commercial AI APIs like OpenAI or Claude, you face three chronic bottlenecks:

• Unpredictable Per-Token Bills: Processing large documents, source code repositories, and high-frequency webhook streams quickly escalates into thousands of dollars each month.
• Data Privacy & Compliance: In healthcare, fintech, legal, and proprietary software engineering, piping confidential customer data or proprietary algorithms into third-party cloud APIs triggers major compliance violations.
• API Rate Limits & Outages: Peak-hour throttling and third-party downtime halt internal automated agents and customer-facing features.

Running on our DeepSeek & AI VPS nodes provides 100% data sovereignty, flat monthly costs, and an air-gapped private inference endpoint.

Step 1: Choosing the Right VPS Sizing for DeepSeek-R1

DeepSeek-R1 models come in several distilled parameter sizes. Because quantized model weights reside entirely in RAM during inference, your server memory dictates which model size you can run:

• DeepSeek-R1-Distill-Qwen-7B (Q4_K_M): Requires 16 GB RAM (4 vCPUs). Fast, agile, and handles code completion, customer service, and structured JSON parsing at ~45 tokens/sec.
• DeepSeek-R1-Distill-Qwen-14B (Q4_K_M): Requires 32 GB RAM (8 vCPUs). The sweet spot for complex reasoning, multi-step agent planning, and in-depth code refactoring.
• DeepSeek-R1-Distill-Qwen-32B (Q4_K_M): Requires 64 GB RAM (16 vCPUs). Near-parity with full-sized frontier models across mathematical and logical benchmarks.

Step 2: Installing Ollama on Ubuntu / Debian

Connect to your VPS via SSH and run the official automated installer script:

curl -fsSL https://ollama.com/install.sh | sh

Once installed, verify that the Ollama service daemon is active and running:

systemctl status ollama

By default, Ollama binds to 127.0.0.1:11434. If you intend to connect external microservices or web applications directly to this server, edit the systemd service file to bind across your private static IPv4:

systemctl edit ollama.service

Add the following environment override block:

[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
Environment="OLLAMA_NUM_PARALLEL=4"

Save and reload the systemd configuration:

systemctl daemon-reload && systemctl restart ollama

Step 3: Pulling and Testing DeepSeek-R1

Now download and run your target DeepSeek-R1 model weight directly inside your terminal:

ollama run deepseek-r1:14b

Our high-speed 1 Gbps datacenter network in Frankfurt or Mumbai will pull the 9 GB model weights in under 30 seconds. You will be greeted by an interactive prompt where you can inspect DeepSeek's <think> step-by-step reasoning tokens live.

Step 4: Using the OpenAI-Compatible API Endpoint

One of Ollama's most powerful native features is its built-in OpenAI API wrapper. You can call your Hostinap AI VPS using standard cURL, Python, or TypeScript:

curl http://YOUR_SERVER_IP:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-r1:14b",
    "messages": [
      {"role": "system", "content": "You are an expert Linux sysadmin assistant."},
      {"role": "user", "content": "Write a bash script to monitor MySQL query latency."}
    ],
    "temperature": 0.6
  }'

This allows you to point IDE extensions like Cursor, Continue.dev, Flowise, or LangChain straight to your own VPS IP address, giving your engineering team unlimited AI-assisted coding with zero per-seat subscription fees.

Step 5: Deploying Open WebUI via Docker Compose

If you want a modern ChatGPT-style web workspace for your team with document upload RAG, user permissions, and chat history, install Open WebUI using Docker Compose:

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

Navigate to http://YOUR_SERVER_IP:3000 in your browser to create your admin account. The dashboard automatically syncs with your local Ollama models.

Summary & Hardware Recommendations

Self-hosting DeepSeek-R1 puts complete control of your AI pipeline back into your hands. For production deployments with continuous multi-agent traffic, we recommend pairing high-frequency AMD EPYC dedicated cores with pure PCIe 4.0 NVMe storage to prevent model memory thrashing.

Explore our pre-configured DeepSeek & AI VPS Hosting plans starting at \$25/mo, where our senior DevOps team handles the complete Docker, Ollama, and SSL reverse proxy configuration for free.