Best VPS for Ollama (2026): Where Cheap Hosting Hits Reality

Ollama is the easiest way to run local LLMs. It also makes it easy to deploy something that does not actually work for your use case, because the install command is the same whether you have 8 GB of RAM or an H100. The framework will not tell you that 2 tokens per second is too slow for a chatbot.

Here is the brutally honest hosting math after testing Ollama on six different VPS configs.

Ollama Is a Wrapper, the Model Is the Cost

Ollama itself is light. Around 100 MB of Go binary and runtime overhead. Every resource decision is about the model: which one you run, in which quantization, and whether you have GPU.

CPU inference works for the install demo and breaks for daily use. The numbers I measured on a Hetzner CCX33 (8 vCPU, 32 GB RAM, no GPU):

• Llama 3.1 8B Q4_K_M: 5 to 7 tokens per second
• Llama 3.1 8B Q8_0: 2 to 3 tokens per second
• Llama 3.1 70B Q4 (when it fits): 0.5 tokens per second

Compare to GPU on a Hetzner GEX44 (RTX 6000 Ada, 48 GB VRAM):

• Llama 3.1 8B Q4: 90 to 120 tokens per second
• Llama 3.1 13B Q4: 60 to 80 tokens per second
• Llama 3.1 70B Q4: 20 to 30 tokens per second
• Llama 3.1 70B Q8: 15 to 20 tokens per second

The CPU path is for batch jobs that you can leave overnight. The GPU path is what you need for anything interactive.

VPS Comparison for Ollama

Hetzner Cloud CCX23: For 7B CPU inference

The minimum I would consider for Ollama with a 7B model. 16 GB RAM fits the model plus context, and dedicated CPU keeps the 5 to 7 tokens per second steady. Use this for personal assistant work, document summarization, or any batch task where 2-second response times are acceptable.

Do not use this for chatbots or interactive applications. The latency makes it painful.

Pros:

• Cheapest realistic Ollama setup
• 16 GB RAM gives headroom for 7B model plus reasonable context
• Dedicated CPU keeps inference rate steady

Get Hetzner: Hetzner Cloud.

Hetzner Cloud CCX33: For 13B CPU dev work

32 GB RAM fits a 13B model with comfortable context. CPU inference at 2 to 3 tokens per second is fine for development and testing, painful for production. Use this when you need to validate that your prompts work on a 13B model before committing to GPU hosting.

Honest take: 59.61 EUR a month for slow CPU inference is hard to justify against renting a GPU hourly via Vast.ai for testing.

Hetzner GEX44: The serious production tier

184 EUR a month for an RTX 6000 Ada with 48 GB VRAM is the cheapest path to production Ollama in Europe. The 48 GB VRAM fits 70B Q4 models with room for context, and 13B Q8 models with very large context (128K+).

This is the right tier when:

• You serve a chatbot or interactive application
• Token cost would exceed 200 USD a month on API aggregators
• You need data residency or air-gapped deployment

The 192 GB system RAM is overkill for Ollama itself but useful when you also run a Flowise or LangGraph orchestrator on the same machine.

Get Hetzner: Hetzner Cloud.

Contabo VPS XL: For massive CPU batch jobs

27.49 EUR a month for 96 GB RAM lets you load 70B Q4 models on CPU. Inference at 0.5 to 1 token per second is unusable interactively but tolerable for overnight batch processing of document corpuses.

Pick this for niche batch use cases where total monthly throughput matters more than per-request latency.

Get Contabo: Contabo VPS.

What I Would Pick

For production Ollama with any interactive use: Hetzner GEX44 with GPU. For development and testing with 7B models: Hetzner CCX23. Skip CPU inference for 13B+ models unless you have a batch use case that genuinely fits. The token-per-second math does not work for anything else.

The broader VPS landscape sits at the SelfHostVPS comparison. Ollama pairs well with OpenWebUI, AnythingLLM, and most agent frameworks. See those guides for combined deployment recommendations.