Ollama for business: what is missing before you run it in production
- Updated
- By
- The MagnolIA team
- On this page
- 7 min read
Short answer
Ollama is an open-source program, under the MIT licence, that downloads and runs language models on your own machine. For one person or a small technical group it is often enough. A whole office also needs what it does not include: users, permissions, quotas, audit, backups and maintenance. A managed service such as MagnolIA includes them.
What is Ollama and how is it licensed?
Ollama is a program that downloads open language models and runs them on your own computer or server. Its documentation sums it up as using open models in your desktop apps and coding agents, or building them into your application. It installs on macOS, Windows and Linux, and exposes its own API, which also accepts clients written for the OpenAI and Anthropic APIs.
Ollama's code is published under the MIT licence, which allows commercial use. That covers the program, not the models: each open-weight model has its own licence, and they do not all allow the same things. Read it before taking a model into production.
Ollama is one of the tools used to build a self-hosted AI server; that guide compares it with others. This one sticks to a single question: it already works on one machine, so what is missing before the whole office can use it?
What does Ollama do well?
- It starts with very little configuration. Install it, pull a model and you can start asking.
- It ships closed to the outside. By default it listens only on the machine itself, at the local loopback address on port 11434.
- When run locally, it does not see your data. Its documentation states that when you run it locally, Ollama does not see your prompts or data.
- It manages memory for you. It loads the model onto the graphics card if it fits, spreads it across several if not, and keeps it in memory for five minutes after the last use.
- It is easy to update. On macOS and Windows it downloads updates by itself; on Linux you re-run the installer.
- It speaks the language of many tools. Because it accepts clients of the most widely used APIs, editors and coding agents connect by changing an address.
For learning, for a single developer or for a small technical group, that is usually enough.
What are its defaults, and what do they mean with several people?
Ollama's defaults are designed for one machine and one person. None of them is a flaw: they are sensible choices for that use, and all of them can be changed. It is worth knowing them before giving more people access.
| Setting | Default | What it means with several people |
|---|---|---|
| Listening address | The machine's own loopback address, port 11434. Changed with OLLAMA_HOST. | Nobody else can reach it until you open it to the network. Once open, anyone who can reach that port can use it. |
| Local API authentication | None required. | There are no users or keys: identity has to be put in front of it. |
| Parallel requests per model | 1 (OLLAMA_NUM_PARALLEL). | The second person waits for the first to finish, unless you raise it. The memory required grows with parallelism and with context size. |
| Request queue | 512 (OLLAMA_MAX_QUEUE). When the queue is full it returns a 503 error. | With many people at once, waits grow long before any error appears. |
| Context window | It depends on GPU memory: 4,000 tokens below 24 GiB, 32,000 between 24 and 48, and 256,000 from 48 upwards. Change it with OLLAMA_CONTEXT_LENGTH. | A long document does not fit in full unless you enlarge it, and enlarging it takes more memory. |
| Models loaded at once | Three per graphics card (OLLAMA_MAX_LOADED_MODELS), if they fit in memory. | If everyone picks a different model, they can push one another out of memory. |
| Cloud features | Available; switched off with OLLAMA_NO_CLOUD=1. | If someone uses a model from Ollama's cloud, that request is processed elsewhere. If data must not leave, switch them off. |
What do you have to add before the whole office can use it?
Seven pieces. None is exotic, but each is work, and someone has to keep it running afterwards.
Identity and encryption
A reverse proxy with TLS in front of Ollama's port, tied to the company directory, or an interface that manages users and roles, such as Open WebUI (whose licence is worth reading before commercial use). That way each person signs in with their own account and access can be withdrawn.
Permissions and quotas
Which models each group may use, and how much. Without a limit, a single automation can occupy the server and leave everyone else waiting.
Audit
Ollama's logs are server logs (on Linux, read with journalctl), and its documentation does not describe a per-user record. Knowing who asked what and when is a job for the layer you put in front.
Availability
The documentation consulted describes one server on one machine. If the service cannot stop, the second machine and the load balancing between the two are built separately.
Controlled updates
On Linux you can pin the version at install time (OLLAMA_VERSION). Test each new version and each new model against your own tasks before switching everyone over.
Backups
Models can be downloaded again. What cannot be recovered is yours: conversations, indexed documents, users and the configuration of the layer in front. That is what needs backing up, with a tested restore.
Monitoring
Graphics card memory, queue length, 503 errors and disk space, with an alert that reaches a person.
The checklist for deciding whether you are in a position to take that on is in building your own AI server.
When is Ollama enough, and when is it not?
| Situation | Is Ollama enough? | Why |
|---|---|---|
| One person on their own computer | Yes. | It is the use it is configured for. |
| A technical team of a few people on a trusted network | Usually, yes. | They tolerate a wait or a failure and know how to fix it. |
| A prototype or an internal trial | Yes. | What matters is learning quickly which model works. |
| The whole office, including non-technical staff | Not on its own. | You need users, permissions, support and someone who answers when it fails. |
| Personal or confidential data | Not on its own. | You have to be able to show who accessed what, and that takes identity and logging. |
| A service that cannot stop | Not on its own. | A single machine with no standby is a single point of failure. |
What does a managed service add on top of a model server?
The same seven pieces, already built and with someone answerable for them. In MagnolIA's case, the software that serves the models (MagnolIA Server) identifies each user, applies their permissions and quotas, spreads the work across machines and logs usage. The audit trail is kept for two years in the installation, backups are encrypted with AES-256, and MagnolIA selects, tests and updates the models.
In exchange there is less control: users have no technical access to the system, and operating it is the provider's job. For a team that enjoys choosing every component, that is a loss. What is not lost is compatibility: MagnolIA Server exposes an API compatible with those of OpenAI and Anthropic, so tools that point at Ollama today connect by changing the address and the key.
What any such service should include, whoever provides it, is in managed on-premise AI. How MagnolIA's is built is in MagnolIA's AI server. And if the reason for keeping the model in-house is not depending on a connection, see AI without an internet connection.
Frequently asked questions about Ollama for business
Can a business use Ollama?
A business can use Ollama: it is software under the MIT licence, which allows commercial use, and it runs open models on your own machine. Each model also has its own licence, which is worth reading. For a small technical group it is often enough; a whole office also needs identity, logging, backups and someone to maintain it.
Does Ollama have user authentication?
Ollama's local API does not require authentication, according to its documentation: anyone who can reach port 11434 can use it. By default it listens only on the machine itself. If you open it to the network with the OLLAMA_HOST variable, identity, encryption and permissions have to go in front of it, with a reverse proxy or a user-management interface.
How many requests can Ollama handle at once?
By default Ollama handles one request at a time per model (the OLLAMA_NUM_PARALLEL variable is set to 1) and queues the rest, up to 512; when the queue is full it returns a 503 error. Parallelism can be raised, but the memory required grows with the number of parallel requests and the context size, according to its documentation.
Does Ollama send my data to the internet?
When Ollama runs locally, its documentation states that it does not see your prompts or data. It also offers models in its own cloud: if those are used, requests are processed off the machine. Those features can be switched off with the OLLAMA_NO_CLOUD=1 variable or in its configuration file, which is advisable if information must not leave the network.
Sources
External figures on this page link to their original source.
Ollama. Official documentation: FAQ. Accessed November 2026. docs.ollama.com/faq
Ollama. Official API documentation: Authentication. Accessed November 2026. docs.ollama.com/api/authentication
Ollama. Official GitHub repository: LICENSE file (MIT License). Accessed November 2026. github.com/ollama/ollama/blob/main/LICENSE
Ollama. Official documentation: Context length (default context size by GPU memory). Accessed November 2026. docs.ollama.com/context-length
Ollama. Official documentation: Troubleshooting (where the server logs are). Accessed November 2026. docs.ollama.com/troubleshooting
Open WebUI. Official documentation: features. Accessed November 2026. docs.openwebui.com/features/
Open WebUI. Official documentation: Open WebUI License. Accessed November 2026. docs.openwebui.com/license/
Is Ollama already working for you? Let's see what is missing
Tell us how many people would use it and with what data. If what you have is enough, we will say so.
Request a pilot


