The appeal is straightforward: host your own AI, keep sensitive data under your own roof, bypass the servers of OpenAI or Anthropic. A widely shared tutorial has made this prospect seem accessible—run Qwen 3.5 9B or Qwen 3.6 35B on Ollama, either locally or on a cheap VPS, and your private model is yours to control. The promise is real. So are the gaps between what’s advertised and what actually works.
The entry price looks forgiving. A Hostinger VPS with 8 vCPU, 32 GB RAM, and 400 GB NVMe is listed at $25.99 per month if you prepay for 24 months upfront—totaling $623.76 paid in advance. Renewal? $49.99 per month. Qwen 3.6, which uses 35 billion total parameters but only 3 billion active per token, requires roughly 20 GB of RAM. It fits the box.
But “private” depends on what you do with that box. By default, Ollama listens on port 11434 and requires no password. The documentation suggests it should bind only to localhost, but security researchers report inconsistencies in its actual behavior. A single misconfiguration—changing one line to adjust OLLAMA_HOST—and your model’s API broadcasts to the internet.
The consequences are concrete. Recent studies documented approximately 175,000 exposed Ollama servers worldwide. In May 2026, researchers disclosed a critical vulnerability, CVE-2026-7482, with a CVSS score of 9.1. A specially crafted request to port 11434 can leak prompts, conversation histories, and environment variables containing API keys. It requires no authentication, and just three API calls trigger the exposure.
Privacy from AI vendors is genuine—OpenAI does not train on your data by default. But you now own the responsibility for securing that data. Hostinger’s hosting agreement explicitly excludes compliance with HIPAA or PCI-DSS. If your data is health-related or financial, this VPS is not the solution.
The tutorial makes it sound simple. For users who understand what they’re doing, it largely is. But for those copying steps, the risks accumulate. Ollama’s default context window is only 4,096 tokens. Local performance depends entirely on your hardware. A quietly misconfigured installation can end up exposed to the world.
This is not an argument against running models locally. It is an argument for going in with eyes open. Your Ollama version needs updating against known vulnerabilities. Qwen 3.5 may have compatibility issues within Ollama that the guide doesn’t mention. And port 11434 must stay closed to the internet, not just firewalled in your mind.
Privacy isn’t a default setting. It’s a series of choices you have to make first.
