Local AI means tasks run on your hardware. Cloud AI means everything runs through an external provider. Most businesses are best served by a hybrid setup: tasks run locally, with language processing handled by a chosen provider. This article compares the options fairly.
Local AI means the AI environment, including projects, knowledge, configuration, and task execution, runs on a system at your location. You control the hardware, access, and data.
At Voltti, the Hermes environment runs locally. That does not mean everything is offline: it can use an external model for language processing. You can read more on our page about local AI for businesses.
Cloud AI means everything runs through an external provider: ChatGPT, Copilot, or another SaaS solution. You send your question to the cloud and receive an answer. The provider determines where your data is processed.
The main differences are where your data is processed, how much control you keep, and how you pay. The hybrid column sits in the middle on purpose: it is the setup most businesses end up with.
| Feature | Local AI | Hybrid | Cloud AI |
|---|---|---|---|
Where data is processed | At your location | Tasks at your location; language processing at the model provider you choose | At the cloud provider |
Control over task execution | High: you set the rules | High for tasks; the provider's terms apply to language processing | Limited to the provider's settings |
Access to your systems | Yes, within your network | Yes, within your network | Only through cloud connections |
Hardware required | Yes, with a GPU for a local language model | Yes, a capable computer | No |
Cost model | Installation + support | Installation + support, plus language processing via API or subscription | Usage-based / subscription |
Suited to | Strict data rules or work that must run offline | Most businesses with recurring work | Occasional questions without links to your systems |
Four questions usually settle it: how sensitive your data is, which model the task needs, whether the work must keep running offline, and how often you use it.
Start with your data. Customer files, quotes, and HR documents are easier to keep under control when files, knowledge, and task execution stay in your own environment. In a hybrid setup, only the context a task needs goes to the model provider you chose, and you decide what that context is.
If personal data is involved, the GDPR also applies in Belgium and the Netherlands. The Dutch Data Protection Authority warns: "Most companies behind the chatbots store all data entered." According to the regulator, you can also arrange with the provider that it does not store the data you enter. Before you connect that data, work through our AI safety checklist for business data.
Then look at the task. Complex reasoning and long documents often work best with a strong cloud model such as Claude. Simpler, repetitive work, such as sorting mail or summarizing short notes, can run on a smaller open-weight model if your hardware allows it. Cloud AI wins here on raw capability, and we say so openly.
How far local AI goes depends on your machine. Our own workstation at Voltti has an RTX 5090 (32 GB of VRAM) and 128 GB of system RAM. It runs four large open-weight models, one at a time. Three of them each have their own role. Qwen3.8-Flash-Next is the default for code and agent work, at about 79 tokens per second. GLM-5.3-Flash is the second opinion, at about 30 tokens per second. We use DeepSeek-V4.1-Flash for hard questions we can wait for: about 25 tokens per second, with roughly 3 seconds before the first token appears. So even locally, you pick the model per task. These are our measurements from October 21 and 22, 2026 on one machine, not a general benchmark. Our article about what computer you need to run AI locally covers the full setup.
Offline work and speed matter less for most offices. They do decide the choice when a connection is not guaranteed, or when a task has to answer within a split second on site.
Finally, count how often you use it. Occasional questions fit a subscription. Recurring work across a team is where a setup of your own starts to pay off, as the next section explains.
That depends on how much you use it. With cloud AI you sign up and can start right away. Local AI needs a one-time installation, on hardware you already have or on a computer you choose for it. For occasional questions, a subscription makes more sense.
For recurring work (quotes, reports, customer service, data analysis), per-user subscriptions add up with every person who needs access. A hybrid setup has a one-time setup cost instead, plus language processing through the API or subscription you choose. Open-weight models also come without per-user license fees: the cost sits in hardware, installation, and support. What a setup costs for your business is always worked out in a custom quote.
Also count what never shows up on an invoice. Your files, knowledge, and task execution stay within your own walls. You do not depend on one provider's price changes. And you can adapt the environment as your processes change.
Our page on security and privacy explains how we handle your business data. Your own setup also needs maintenance. Your team can run updates itself, since Hermes is open source, or you can leave updates and monitoring to us through support. And it needs a plan for failure: backups go encrypted to a separate location in your network, never onto the same machine. Our article on what happens to your AI environment if hardware fails walks through recovery.
The hardware itself does not have to be extreme: a capable computer is enough for the standard setup. You only need heavy hardware if you also want to run the language model locally, and how heavy depends on the model. Should your team get a separate, managed Linux computer? That depends on how many people use the agent, as we explain in when you need a Linux computer for an AI agent.
In practice, most businesses are best served by a hybrid setup. The AI environment (projects, knowledge, configuration, and task execution) runs at your location. For language processing, you use an external model provider of your choosing.
Microsoft describes the same route in its developer guide on cloud and local AI models: "Many production apps use a hybrid strategy: try a local Windows AI API or local model first, then fall back to a cloud endpoint" when, for example, the task requires a larger model. In other words: local where it works, the cloud where a task needs more.
With Hermes you make that split per kind of work. A task that reads internal documents can use a model you trust with that data, or a local open-weight model. A task that drafts a public blog post can use a strong cloud model. What leaves your environment is decided per task, by you.
It also works the other way around. A cloud model can stop in the middle of a task for reasons you do not control: a safety filter, an outage, or a rate limit. That happened to us on October 22, 2026, in one of our own Hermes sessions. We switched to the local Qwen model in the same session: same tools, same session history, only the model changed. It was slower, with some steps taking 70 to 115 seconds, but the work kept going. So a local model does not replace a strong cloud model. It is insurance for when that cloud model is not available. Our article on a local model as a fallback for your AI agent shows how to set one up.
Our article on what stays local in a hybrid AI setup shows which parts stay local and which go to a model provider. A small real-world example is our home agent Sisu on a Raspberry Pi: the base runs at home, while the language model uses an external service.
Choose cloud AI if you only ask occasional questions and do not need an integrated work environment. Choose local or hybrid AI if you want the agent to work with your projects, use your tools, and operate within your boundaries.
Read more about the difference between an agent and a chatbot in our article on AI agents vs. chatbots.
No. Claude is a closed model from Anthropic and is only available through the cloud. Running models locally requires open-weight models, such as Qwen, GLM, or DeepSeek on Hugging Face. In a hybrid setup you combine both. Task execution runs on your hardware, and for language processing you pick a suitable model per task: a local open-weight model or a cloud model like Claude.
Not necessarily. At Voltti, the Hermes environment runs locally, but language processing can use an external model provider. That is called hybrid AI. Without an internet connection, the agent can still reach your local files, but not that provider. A fully offline setup is possible: it needs a local open-weight model and a computer with enough GPU power to run it.
Not necessarily, but you hand over some control. Whatever you send to a cloud service is processed under that provider's terms. If you work with personal data, that processing must also comply with the GDPR. With local or hybrid AI, tasks run on your own hardware and you decide what goes to a model provider. Before you connect business data, agree which data the agent may send and to which provider.
Yes. With Hermes Agent you choose the model provider, so you can start with a cloud model and switch later. If your use case and hardware justify it, we add a local open-weight model. Your projects, knowledge, and configuration stay where they are, so a switch does not mean starting over.
Not necessarily a server. For the standard setup, a capable computer is enough, and that can be hardware you already have. A GPU only becomes necessary when you want to run a language model locally. A separate computer makes sense when a whole team uses the agent or when tasks must keep running after laptops close.
