Choosing the Right AI Model for Your Private Deployment

One of the most important decisions in a RowSpeak Private Deployment is which AI model to use. The right choice depends on your organization's privacy requirements, infrastructure, performance needs, and internal policy.

This guide walks through the options and helps you decide.

The Two Paths

Decision Framework

Use this to narrow down your choice quickly.

Choose open-source if:

• Your data cannot leave your network under any circumstances
• You are deploying in an air-gapped environment
• You have regulatory or compliance requirements around data residency
• You want to avoid dependency on third-party API availability
• You want predictable costs without per-token pricing

Choose closed-source if:

• Your data sensitivity allows external API calls (with your own account)
• You want the highest possible output quality for complex reasoning tasks
• You do not have GPU infrastructure available
• You want to get started quickly without model setup
• You are running a pilot before committing to GPU hardware

Mix both if:

• Different workflows have different sensitivity levels
• You want open-source for sensitive data and closed-source for non-sensitive tasks
• You want a fallback model if one is unavailable

RowSpeak supports routing different workflows to different models within the same deployment.

Open-Source Model Comparison

All open-source models run locally. No internet connection required after initial setup. Licenses allow commercial use.

Closed-Source Model Comparison

You use your own API key. RowSpeak does not act as an intermediary — your data goes directly from your environment to the provider.

What Happens to Your Data with Closed-Source Models

When you use a closed-source model, your spreadsheet data is sent to that provider's API for processing. This means:

• The provider processes your data according to their terms of service
• Enterprise API agreements with OpenAI, Anthropic, and Google typically include data processing agreements (DPAs) that restrict training use
• You should review the provider's data handling policies before enabling this option
• RowSpeak recommends using closed-source models only for non-sensitive data, or after reviewing and accepting the provider's enterprise data terms

For maximum data sovereignty, use open-source models.

Model Routing in RowSpeak

RowSpeak supports configuring different models for different use cases within the same deployment.

Example configuration:

Workflow: Financial reports → DeepSeek V3 (local, sensitive data)
Workflow: Marketing summaries → GPT-5.4 (API, non-sensitive)
Workflow: Default → Qwen3.5-35B (local, general use)
Copy

This lets your organization apply the right model to each workflow based on data sensitivity, without forcing a single choice across all use cases.

Frequently Asked Questions

Can I switch models after deployment?
Yes. Model selection is a configuration change, not a re-deployment. Your IT team can update the model routing configuration without downtime.

Do I need to download model weights myself?
No. The RowSpeak Deployment Pack includes guidance on model acquisition. For air-gapped environments, we provide instructions for pre-loading model weights before deployment.

What if I want to use a model not on this list?
Contact us. RowSpeak's model layer is designed to be extensible. If you have a specific model requirement, we can discuss compatibility.

Can I use a fine-tuned or custom model?
This is available on the Enterprise tier. Contact us to discuss your requirements.

Need Help Deciding?

Book a demo and we will help you map the right model strategy to your environment, data sensitivity, and performance requirements.

You can also review the technical architecture document for more detail on how the model layer integrates with the rest of the system.