The Ollama library holds two kinds of models. Models with a size such as 8b, 30b or 120b are downloaded and run on your own machine. Models tagged "cloud" run on Ollama's servers: you sign in, use up usage credits, and your input leaves your network. Of the 50 newest entries in the library, fourteen were cloud-only on November 15, 2026, among them Mistral Large 4, Muse Spark Open, DeepSeek V4.1 Flash, Kimi K3 and GLM-5.3. According to Ollama, cloud requests are processed mainly in the United States. If nothing should leave your network, switch the cloud features off with OLLAMA_NO_CLOUD=1. Which local model suits you depends mostly on your memory: with 16 GB, models like gpt-oss:20b run; with 24 to 32 GB, the strong models of the 30B class; and from 80 GB, models like gpt-oss:120b.
I run my business with a team of fifteen AI employees built on Claude and test local models on a DGX Spark. Ollama is installed there, but my main model runs under SGLang. For a quick start on a Mac or a PC, Ollama is still the tool I recommend first. Everything I write about local AI is collected on my AI page.
Local or cloud: how to tell the difference
Ollama's docs say about the cloud: "Run larger models in Ollama’s cloud without requiring a powerful GPU. Inference runs on Ollama’s cloud service." In the library you can spot a cloud model by three things:
- In the list, the model carries the tag "cloud" and no size such as 8b or 30b.
- On its tags page, instead of a file size such as "18GB", it shows a usage level such as "Low Usage", "Medium Usage", "High Usage" or "Extra High Usage".
- The tag is called
cloud or ends in -cloud, for example gemma4:cloud or gpt-oss:120b-cloud.
Some models come both ways. You can download Gemma 4 and gpt-oss in several sizes, and both also have cloud tags. Whether a call runs locally or in the cloud then depends on the tag you start.
The fourteen cloud-only models among the 50 newest entries on November 15, 2026 were: Inkling 2 Small, Kimi K3.1 Pro, MiniMax M3.1 Flash, Beam, Muse Spark Open, Mistral Large 4, MiMo V2.6 Pro, MiMo V2.6 Flash, DeepSeek V4.1 Flash, GLM-5.3 Flash, GLM-5.3, Kimi K3, GLM-5.2 and Kimi K2.7 Code. According to their model cards and Ollama's descriptions, most of them have several hundred billion to around three trillion parameters: Kimi K3 has about 2.8 trillion, and Ollama places Kimi K3.1 Pro in the three-trillion class. Models like that fit on no normal machine, which is why Ollama only offers them on its own servers. For Mistral Large 4, the weights have been openly available on Hugging Face since October 29, 2026, yet in Ollama the model is still cloud only.
What happens to your data
For local use, Ollama's statement is clear: "Ollama runs locally. We don’t see your prompts or data when you run locally." For cloud models, Ollama's privacy policy (last updated March 2026) says that prompts and responses are processed only transiently to answer the request and are not used for training. On the pricing page it says about location: "Ollama hosts models and compute resources primarily in the United States. To serve global demand, we may route to Europe for additional capacity." The models run there with partners from Nvidia's cloud provider program, which, according to Ollama, it requires to keep no logs and store nothing.
For you this means: as soon as you use a cloud model, your text usually goes to the United States, sometimes to Europe, even if it is not supposed to be stored there. For transfers abroad, Ollama promises "appropriate safeguards" where the law requires them. I did not find a data processing agreement on Ollama's privacy and pricing pages. For client or employee data you need a legal basis and, as a rule, exactly such an agreement. I'm a developer, not a lawyer. This is not legal advice, just the operator's perspective. How I sort this out for my own setup is described in Local AI and Privacy: Cloud or Server?.
If you want to be sure that nothing leaves your network, turn the cloud features off completely. According to Ollama's FAQ, you do that in the file ~/.ollama/server.json:
{
"disable_ollama_cloud": true
}
Or with an environment variable:
OLLAMA_NO_CLOUD=1
The variable has to apply to the Ollama service, not just your terminal. On a Mac, according to Ollama, you set it with launchctl setenv OLLAMA_NO_CLOUD 1 and restart the app; on Linux you add it to the service with systemctl edit ollama.service as Environment="OLLAMA_NO_CLOUD=1". The entry in server.json is usually simpler. After a restart, Ollama's log shows "Ollama cloud disabled: true". The cloud models and Ollama's web search go away; local models keep running.
What the cloud costs
Locally, Ollama costs nothing; the pricing page says: "Running models on your own hardware is always unlimited." For the cloud, the pricing page lists five tiers, as of November 15, 2026:
Anything beyond the credits is billed per million tokens. For input and output, the pricing page lists for example 0.15 and 0.60 US dollars for gpt-oss:120b, 0.70 and 2.10 US dollars for Mistral Large 4, and 3.00 and 15.00 US dollars for Kimi K3. On weekdays outside 12:00 to 18:00 UTC and at weekends, some models have lower off-peak prices.
Which Ollama model fits your memory
The file size is not the whole memory requirement. The context needs extra room, and according to its docs Ollama sets it based on GPU memory: 4k tokens below 24 GiB, 32k from 24 to 48 GiB and 256k from 48 GiB. For agents and coding tools Ollama recommends at least 64,000 tokens. So add a reserve to the file size. The grouping below is my rule of thumb; I have not measured it.
For gpt-oss-20b, OpenAI itself writes that the model runs "within 16GB of memory"; for gpt-oss-120b, OpenAI names a single 80 GB GPU. For Muse Glimmer, Meta names 24 or 32 GB. The other assignments follow from the file size plus a reserve.
How to install Qwen 3.8 in Ollama step by step, which tag fits which memory and which two settings you should change afterwards is covered in my own guide Installing Qwen 3.8 with Ollama. This article is about choosing across all models. How the three strongest models of the 30B class differ is compared in Muse Glimmer vs Qwen 3.8 vs Gemma 4. I cover the Mistral models in the table in detail in Running Mistral Locally: Models and Licenses.
The license is part of the choice
A model that runs locally in Ollama is not automatically yours to use for everything. Most models in the table are under Apache 2.0 and so free for commercial use. The Nemotron models are under OpenMDW-1.1; for the Super version Nvidia writes: "This model is ready for commercial or non-commercial use." Mistral Medium 3.5 excludes companies with more than 20 million US dollars in monthly revenue. And Qwen3.8-Flash-Next, which Ollama lists as a local model at 105 GB, has its own license with special rules for providers of AI services. So read the license on the model card before you use a model for clients. This is not legal advice. The key clauses side by side are in my article on open-weight licenses.
When Ollama is no longer enough
Ollama is built for one user on one machine. As soon as several people, an agent and a script access the same model at the same time, or a model needs to be reachable around the clock, I switch to a serving framework such as SGLang or vLLM. On my DGX Spark, Qwen3.8-27B runs under SGLang, measured at 50.7 tokens per second in August 2026; on German prose my November test gave 18.2. The path from Ollama to there is described in Setting Up Local AI: Ollama to SGLang. Which models run how fast on the DGX Spark is what I measure in Best LLMs for the DGX Spark, Measured.
Frequently asked questions
Which Ollama models run locally?
All models whose tags page shows a file size in GB. The library list doesn't always show the size; Qwen3.8-Flash-Next, for example, runs locally even though the list names no size. Models tagged only "cloud" run exclusively on Ollama's servers.
What are Ollama cloud models?
Models that Ollama runs on its own servers, usually because they are too large for a normal machine. You call them from the app, the terminal or the API without downloading the model, but you need an account and usage credits.
Are Ollama cloud models GDPR compliant?
That depends on your use and isn't something I can answer across the board. According to Ollama, processing happens mainly in the United States, input is not used for training and is not supposed to be stored. I did not find a data processing agreement on its pages. For personal data, use local models or have it checked. This is not legal advice.
How do I turn off the cloud in Ollama?
Add "disable_ollama_cloud": true to ~/.ollama/server.json or set OLLAMA_NO_CLOUD=1 for the Ollama service, on a Mac with launchctl setenv, on Linux in the service configuration, and restart Ollama. The log then shows "Ollama cloud disabled: true".
What does Ollama cost?
Locally, nothing. For the cloud there is Pay as you go with no base fee, where you pay per token from purchased credits, Pro at 20 US dollars and Max at 100 US dollars per month, as of November 15, 2026. Anything beyond the included credits is billed per million tokens.
Which Ollama model fits into 16 GB?
gpt-oss:20b at 14 GB, for which OpenAI explicitly names 16 GB, plus Devstral Small 2 at 15 GB for coding and Ministral 3 14B at 9.1 GB. Always add some memory for the context on top of the model.
Where to go from here
If you haven't installed Ollama yet, start with Setting Up Local AI: Ollama to SGLang. If you're planning to use a Mac for it, Local AI on a Mac: Mini as AI Server will help. Before you put client data into a model, read Local AI and Privacy: Cloud or Server?, and for an overview of all open models, see my comparison of open source AI models.
