The choice between Ollama Cloud and a self-hosted Ollama on a rented GPU comes down to two different bills for the same software: a monthly subscription with usage credits on one side, and an hourly GPU rate on the other. Facts below are from ollama.com's cloud page and pricing page, checked 28 September 2026.
What Ollama Cloud costs
Ollama Cloud is a subscription with usage credits on top, not a per-token pay-as-you-go plan on its own. The Free plan is 0 dollars a month with a starter amount of credits and 1 concurrent request. Pro is 20 dollars a month, or 200 dollars a year billed annually, with 60 dollars of usage credits included each month and 3 concurrent requests. Max is 100 dollars a month with 300 dollars of credits and 10 concurrent requests. Team is 500 dollars a month with 1,000 dollars of credits shared across the whole team and 10 concurrent requests. Enterprise pricing is custom. Unused credits do not roll over to the next month, and off-peak pricing applies outside 12:00 to 18:00 UTC on weekdays and all day on weekends.
Which models are cloud-only
Ollama's library carries local, pullable tags for gpt-oss, the current Qwen models, Gemma 4, Llama and several others, which means anyone with the right GPU can run them without a subscription. DeepSeek V4, GLM-5.3, Kimi K3 and MiniMax's models are cloud-only in Ollama's library: there is no local tag to pull. They are also far larger than one GPU: DeepSeek-V4-Pro's checkpoint is close to 893 GB and GLM-5.3's about 756 GB, so running either yourself takes an 8-GPU node of the largest cards and a serving engine other than Ollama. Kimi K3 goes further still: a 2.8 trillion parameter mixture-of-experts model with only 104 billion active per token, and a native checkpoint of about 1,561 GB, past what even an 8-GPU node holds. Cloud-only in Ollama's own library does not mean impossible to self-host anywhere, only that Ollama itself does not publish a local tag most people could ever load.
Some models that do fit one GPU still get a cloud option alongside the local one. gpt-oss and Gemma 4 both appear on Ollama's cloud pricing table as well as in the local library, tags such as gpt-oss:120b-cloud running on Ollama's own hardware instead of yours. That is a choice between paying by usage credit for a model you could also self-host, which is a different question from the DeepSeek or GLM case, where self-hosting through Ollama specifically is not an option at all.
Self-hosting on a rented GPU
The Open WebUI and Ollama template runs the same Ollama engine on a GPU you rent by the hour rather than a subscription with usage credits, and the prompts and model both stay on a VM that only your account can open. It fits the models covered on the models hub: gpt-oss, the current Qwen family and Gemma 4 all run on a single rented GPU, sized by weights plus the KV cache rather than by a credit budget. What it does not do is reach the largest cloud-only models. DeepSeek V4 and GLM-5.3 need a multi-GPU node bigger than most people want to rent for a single chat session, so that is where Ollama's own cloud-only tags are the more practical answer today, if those specific models matter to you. My own line is simple: if a model fits one GPU, I rent the GPU. Ollama Cloud earns its subscription only once a model too large for anything I would rent enters the picture.
The bill for self-hosting works differently too. It is the GPU's hourly rate for however long the instance runs, shown in full on the deploy page before you pay, rather than a monthly credit balance that resets whether or not you used it. A quick test costs a fraction of an hour, and a model left running overnight costs every one of those hours, since DeployGPU has no auto-stop timer to catch a forgotten deployment.
Opens the deploy page on the cheapest 1x 48 GB GPUs offer at 2026-10-16 01:15 UTC: QuantaCloud, $0.48/hr. You can pick another GPU there before paying.
Decision table
| Situation | Better fit |
|---|---|
| Occasional use of a model that fits one GPU, no interest in managing a server | Ollama Cloud's Free or Pro plan |
| Heavy, sustained use of a model that fits one GPU | A rented GPU, billed by the hour rather than by usage credit |
| Prompts and outputs must stay off a third party's servers | A rented GPU running the model yourself |
| Need DeepSeek V4, GLM-5.3, Kimi K3 or a MiniMax model specifically | Ollama Cloud, since none of these has a local Ollama tag |
| Building an app that calls the model many times a day | Either, depending on whether the usage credits or the hourly GPU rate work out cheaper for that volume |
Where this fits on DeployGPU
DeployGPU rents the GPU. It has no relationship with Ollama's own cloud product, and the Open WebUI and Ollama template is a way to run Ollama yourself rather than a way to reach Ollama Cloud. Anyone who needs the cloud-only frontier models has to use Ollama's own service for that part, and can still self-host everything else on a rented GPU here.
Frequently asked questions
Is Ollama Cloud the same company as DeployGPU?
No. Ollama Cloud is a service from the Ollama project itself. DeployGPU rents GPUs, including ones that run Ollama through the Open WebUI and Ollama template, but has no connection to Ollama's own cloud product.
Can I self-host DeepSeek V4 or GLM-5.3 on a rented GPU instead of using Ollama Cloud?
Not through Ollama, which does not publish a local tag for either. Serving them at all needs a large multi-GPU node and a different serving engine than Ollama, which is a bigger commitment than most self-hosting use cases call for.
Does self-hosting mean my prompts never leave my own computer?
It means they stay on a VM you control rather than a third party's inference servers. The GPU itself is still rented hardware from a cloud provider, so treat it like any cloud server rather than a machine under your desk.
Why would anyone pick Ollama Cloud for a model that also runs locally?
Convenience and avoiding the GPU altogether. Some people would rather pay for usage credits on demand than manage a server, even for a model small enough to self-host.
Do concurrent request limits matter for self-hosting too?
In a different way. A rented GPU has no concurrent request limit set by a plan, but the GPU's own memory caps how many requests actually run at once, covered on the KV cache explainer.