How to Self-Host DeepSeek R1 on Contabo (2026): Which Model Fits Which Cloud VDS, the Free DeepSeek App, and What a CPU-Only Server Can Really Do

Contabo sells “DeepSeek Hosting”: run DeepSeek R1 on your own server, “with no GPU required”, on one of its Cloud VDS plans (virtual dedicated servers, now marketed as “Max Performance VPS”). In Contabo’s price data there is even a free DeepSeek app for those plans. This guide sorts out which DeepSeek model fits which plan, what the server really costs, how to install it, and the part the sales page glosses over: on a CPU-only server you run the smaller, distilled R1 models, and Contabo’s own VDS page says production-scale AI should still go to a hosted API.

What you actually get: R1 “distilled” models running in Ollama

• Ollama does the work. Contabo’s install guide sets up Ollama, a free model runner, and pulls a DeepSeek R1 model from Ollama’s library. Everything runs on the server’s CPU and RAM.
• The small sizes are distilled models. Ollama’s deepseek-r1:1.5b to :70b tags are DeepSeek-R1-Distill models: per DeepSeek’s model card they are “fine-tuned based on open-source models, using samples generated by DeepSeek-R1” (Qwen 2.5 and Llama bases). The full DeepSeek-R1 is the 671b tag, a 404 GB download, which does not fit in any Cloud VDS (the largest has 96 GB of RAM).
• The distills are still strong for their size. DeepSeek’s own benchmark claim: “DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI-o1-mini across various benchmarks”. That is DeepSeek’s figure, not our test.
• Licences. DeepSeek-R1’s weights are MIT-licensed. Per Ollama’s page, the Qwen 2.5-based distills inherit Apache 2.0 origins and the 70B distill from Llama 3.3 (Llama 3.3 licence), so check those terms for commercial use.
• R1 is a “thinking” model. Ollama tags it “thinking”: it writes out its reasoning before the answer. Our take: that means many more tokens to generate per reply, which you feel most on a CPU.

Which Contabo plan for which DeepSeek model?

Contabo’s DeepSeek page shows four Cloud VDS plans, all with dedicated AMD EPYC 7282 cores and NVMe storage. Prices are in EUR before VAT. The monthly-term price is what you pay on a 1-month contract; Contabo gives 15% off on 12 months and 20% off on 24 months, and the cards show the 24-month rate “for the first 24 months”.

Contabo’s install guide matches models to plans like this (sizes from Ollama’s library; the RAM figure is Contabo’s, based on its rule to “allocate twice the model size in RAM”):

Our reading of those numbers:

• 14b, the model Contabo picks for its own image, fits Cloud VDS M. By Contabo’s 2x rule it needs 20 GB, and Cloud VDS M has 32 GB. The DeepSeek page agrees (“we recommend VDS M and up” for deepseek-r1:14b); only the blog table says VDS L. So €62.00 a month (€49.60 on 24 months) is the realistic entry point for 14b.
• 32b needs about 40 GB by the same rule: Cloud VDS L (48 GB) clears it, though the blog says XL. 70b (43 GB download, 85 GB by the rule) only fits Cloud VDS XXL (96 GB).
• Context costs RAM too. Ollama’s docs: “Setting a larger context length will increase the amount of memory required to run a model”, and with parallel requests “Required RAM will scale by OLLAMA_NUM_PARALLEL * OLLAMA_CONTEXT_LENGTH“. Leave headroom if you plan long documents or several users.
• Speed is the unknown. Contabo publishes no speed figures for DeepSeek on these plans and we have not measured them. Our suggestion: on a monthly term, start with deepseek-r1:7b or :8b to see whether responses are fast enough for you, then try 14b.

Cheaper option: Cloud VDS S (€49.00/mo or €39.20 on 24 months; 6 dedicated vCPU, 24 GB RAM, 180 GB NVMe) isn’t on the DeepSeek page and doesn’t get the DeepSeek app in Contabo’s price data, but 24 GB is above the 20 GB Contabo gives for 14b, and the manual install below works on it. Contabo’s blog warns that plans below VDS M “might be usable, but will definitely affect the LLM performance”.

What you actually pay for the server

• Term: the minimum contract is 1 month, with no setup fee on these plans. On Cloud VDS M a 24-month term brings €62.00 down to €49.60 a month: €1,190.40 over two years against €1,488.00 on monthly terms (our math).
• VAT: prices above exclude VAT. Contabo’s country list shows, for example, 19% for Germany, 20% for the UK and 21% for Romania.
• Location: the prices above are for the EU location. Other regions add a monthly surcharge, and on VDS it is large: on Cloud VDS M from €14.30 (UK or US Central) to €35.45 (Japan); on Cloud VDS L from €20.10 to €49.75; on XL from €27.25 to €67.50; on XXL from €40.20 to €99.55.
• Backups: Contabo’s price data for these VDS plans has no Auto Backup add-on (the Cloud VPS plans have one); it offers FTP backup space instead, from €4.59/mo for 100 GB. Models can be re-downloaded, but back up your chats, prompts and settings.
• No per-token bill. This is the one place where Contabo’s “Pay one monthly fee” pitch is fully true: a model running in Ollama on your server has no usage charge. You pay the server, whether you use it or not.

See Contabo Cloud VDS plans for DeepSeek

What a CPU-only server can really do

The DeepSeek page is enthusiastic: “No GPU required – DeepSeek is optimized for CPU only processing”, models up to 70b, ideas from code review to IoT maintenance. Contabo’s own VDS product page is more careful about the same servers:

“Lightweight models like Gemma3:1b can run on the VDS itself for background tasks. Production-scale inference points at a hosted LLM API.”

That page also suggests “a small local model like Gemma 2B or 4B for background classification or routing” and, “For real inference at scale”, pointing your stack at OpenAI, Anthropic, Mistral or Gemini. Our take: both pages are describing the same hardware, so plan for the careful version.

• Good fits: a private chatbot or document helper for yourself or a small team, testing prompts and RAG pipelines on data you don’t want to send to an API, batch jobs that can run overnight.
• Poor fits: many simultaneous users, real-time chat for customers, coding agents that need long context. Ollama’s docs say agents and coding tools “should be set to at least 64000 tokens”, far above its 4,096-token default, and each step up costs RAM.
• “128,000 tokens at once”: the models support a 128K window (Ollama lists 128K for the distills), but Ollama starts at 4,096 tokens unless you raise it (its context-length page gives 4k below 24 GiB of VRAM, and a CPU-only server has no GPU memory). You set it with OLLAMA_CONTEXT_LENGTH.

How to install DeepSeek on a Contabo VDS

The free DeepSeek app. In Contabo’s price data a “DeepSeek” app is listed at €0 on Cloud VDS M, L, XL and XXL (not on Cloud VDS S, and not on the Cloud VPS plans we checked). The DeepSeek page says deepseek-r1:14b is “the one chosen for our custom image”. However, Contabo’s help center 1-Click Apps folder (10 articles) has no DeepSeek article, and the guide the page links to (its URL still says “1-click image”) now only describes a manual install. If you pick the app at checkout, expect a preinstalled 14b setup, but there is no official how-to for it; after the first SSH login, ollama list shows which model is installed. Contabo’s help center notes that “Only one application can be installed per server”.

Manual install (works on any plan, and lets you pick the model). Contabo’s steps, on Ubuntu 22.04 LTS:

sudo apt update && sudo apt upgrade -y
# install Ollama
curl -fsSL https://ollama.com/install.sh | sh
sudo systemctl start ollama
# download and run the model (14b = ~9 GB)
ollama pull deepseek-r1:14b
ollama run deepseek-r1:14b
# check what is installed
ollama list

Two useful checks from Ollama’s FAQ: ollama ps shows what is loaded, and “100% CPU means the model was loaded entirely in system memory” (expected on a VDS). Models stay in memory for 5 minutes after the last request by default, so the first reply after a pause is slower while the model loads.

Keep it private

• Leave the API on localhost. “Ollama binds 127.0.0.1 port 11434 by default.” Don’t change OLLAMA_HOST to 0.0.0.0 on a server with a public IP unless something with a login sits in front of it. To use it from your computer, open an SSH tunnel:

ssh -N -L 11434:localhost:11434 root@<your-server-ip>
# then call http://localhost:11434 on your own computer

• SSH and firewall. Our suggestion: add an SSH key, turn off password logins, and allow only port 22 (Contabo’s Firewall add-on is listed at €2.49/mo on VDS, or use the server’s own firewall).
• “100% offline” is a promise about the model, not the server. Once a model is downloaded, Ollama answers without calling out to anyone, which is the privacy win. The server itself is still on the internet, so updates and SSH hygiene are on you: Contabo’s terms say maintenance of the included open-source software is “within the end-users responsibility”.

Order a Contabo Cloud VDS for DeepSeek

Want an AI agent instead of a chatbot? Ollama’s DeepSeek page lists one-line launchers for agents such as Hermes Agent and OpenClaw. See Hermes Agent on Contabo and OpenClaw on Contabo (both free 1-click apps on cheaper Cloud VPS plans, usually paired with a hosted model because of the 64K-context need), or Paperclip AI on Contabo to run a team of agents. All Contabo tiers side by side: our Contabo VPS price guide. Or start from Contabo’s homepage.

FAQ

Can DeepSeek run on Contabo without a GPU?
Yes, the distilled DeepSeek R1 models (1.5b to 70b) run on CPU and RAM in Ollama. Contabo publishes no speed figures, and its own VDS page reserves local models for light background tasks, so test with a small model on a monthly term first.

Which Contabo plan do I need for DeepSeek R1 14b?
Cloud VDS M (8 dedicated vCPU, 32 GB RAM, 240 GB NVMe): €62.00 per month on a monthly term, €52.70 on 12 months or €49.60 on 24 months, before VAT. Contabo’s 2x-RAM rule gives 20 GB for 14b; its blog table says VDS L (€87.00/mo).

Can I run the full DeepSeek-R1 (671B) on Contabo?
Not on a Cloud VDS. Ollama lists the full model as a 404 GB download, and the largest Cloud VDS (XXL) has 96 GB of RAM and 720 GB of NVMe.

Is the DeepSeek app on Contabo free?
It is listed at €0 on Cloud VDS M to XXL, and the models are free to download. You pay only for the server. Note the licences: R1 is MIT, the distills inherit Qwen 2.5 or Llama terms.

Do I pay per token when self-hosting DeepSeek?
No. A model running in Ollama on your own server has no usage fee; the cost is the flat monthly server price (plus VAT and any location surcharge).

How do I connect to DeepSeek on my Contabo server from my computer?
Ollama listens on 127.0.0.1:11434. Run ssh -N -L 11434:localhost:11434 root@<your-server-ip> and use http://localhost:11434 locally, instead of exposing the port to the internet.

Does Contabo support DeepSeek?
Contabo supports the server (network, hardware) and the initial setup of its 1-click add-ons; the software itself, its updates and troubleshooting are your responsibility, per its terms. Use Ollama’s documentation for the model side.

Sources

All checked on October 22, 2026 (Europe/Bucharest).

• Contabo – DeepSeek Hosting: https://contabo.com/en/hosted-deepseek-ai-enterprise-cloud/
• Contabo – Cloud VDS / Max Performance VPS: https://contabo.com/en/vds/
• Contabo blog – How to Install DeepSeek AI on Contabo Servers (last updated July 8, 2026): https://contabo.com/blog/how-to-install-deepseek-ai-on-contabo-1-click-image/
• Contabo Help Center – 1-Click Apps: https://help.contabo.com/en/support/solutions/folders/103000644094
• Contabo Help Center – What is Ollama Phi4 and how do I use it on Contabo?: https://help.contabo.com/en/support/solutions/articles/103000399425
• Ollama – deepseek-r1 model page and tags: https://ollama.com/library/deepseek-r1/tags
• Ollama docs – FAQ: https://docs.ollama.com/faq
• Ollama docs – Context length: https://docs.ollama.com/context-length
• DeepSeek-R1 model card (Hugging Face): https://huggingface.co/deepseek-ai/DeepSeek-R1