A self-hosted open model suits security work where data must stay in your environment and the task is narrow: classifying logs, summarizing internal tickets, drafting queries. It fits less well when a task needs the strongest available reasoning or a huge context. The limit is your hardware, not a plan.
How is this different from a plan?
With a hosted plan you pay a price and receive a usage allowance. With a self-hosted model there is no allowance. Ollama's pricing page, fetched 18 October 2026, says "Running models on your own hardware is always unlimited" and lists a Free plan at 0 dollars that includes running models locally. The same page lists optional cloud plans, such as Pro at 20 dollars a month or 200 dollars a year, with included usage credits. Those cloud plans are not self-hosting; the page notes that prompt or response data is never logged or trained on for cloud models, a statement about that service only.
The real costs of self-hosting are different ones: hardware or rented GPUs, power, setup, updates and the time of whoever runs it. We did not find a published, comparable figure for those, so we give none. Your own quote from a hardware vendor or cloud provider is the number to use.
Which security tasks fit?
Data that cannot leave. Source code under export restrictions, incident notes, customer logs. A local model avoids sending them to an external service, though you still own access control and logging on the host.
Narrow, repetitive classification. Tagging log lines, grouping similar findings, labeling tickets by team. Small models do this adequately, and it is easy to measure accuracy against a labeled sample.
Drafting and summarizing. Turning a scanner finding into a ticket, summarizing a long incident timeline, writing the first draft of a detection query. A person reviews the result.
Offline environments. Air-gapped networks where no hosted service is reachable.
Which tasks fit less well?
Deep multi-step reasoning across a large codebase, and tasks that need long context, depend on model size and memory. A model that fits on your hardware may not match a hosted frontier model on those, so test it on your own examples rather than trusting a leaderboard. Any model that can call tools also needs the controls OWASP describes for excessive agency: narrow tools, scoped permissions and human approval for high-impact actions.
What should you check before you commit?
Check the model's license for your use. Check how you will update it and track vulnerabilities in the serving software, because the runtime is software with dependencies like any other. Decide who may query it and what is logged. Measure quality on twenty real tasks and write down the failures.
How does Safeguard relate?
Safeguard supports Full AI, Zero AI or bring your own model. Zero AI runs without any model, so a team can start with scanning and reachability analysis and add a model later. Which models and hosting setups are supported for bring your own model is a question for the Safeguard team and its documentation.
What to do
- List security tasks that need data to stay in your environment.
- Mark each as narrow and repetitive, or open-ended reasoning.
- Build a 20 task test set from your own work with known good answers.
- Run candidate local models against it and record failures.
- Check each model's license and the serving software's vulnerability history.
- Restrict who can query it, log use and scope any tool access.
- Keep a hosted option for tasks the local model cannot handle, with clear rules on what data may go there.
Quick check
Is local model use metered?
Not by Ollama. Its pricing page says running models on your own hardware is always unlimited. You still pay for hardware, power and administration.
Does a local model remove the need for review?
No. Output can be wrong or manipulated through its inputs, so treat it as a draft and limit what actions it can take.
Sources
- Ollama pricing (fetched 18 October 2026)
- OWASP GenAI: LLM06:2025 Excessive Agency
Last reviewed: 18 October 2026.
