Ollama Cloud: pricing, reviews, and limitations (2026)

#What Ollama Cloud is

Since 2025, Ollama has offered a paid service called Ollama Cloud, which lets you call models hosted on its GPU infrastructure while using exactly the same commands as for a local model. The promise: run massive open-weight models (hundreds of billions of parameters) without needing a Mac Studio Ultra or an H100 cluster.

In practice, you install Ollama as usual, authenticate, and pull a model marked "cloud" (for example, gpt-oss:120b-cloud). Inference no longer runs on your GPU, but on Ollama's servers—your machine only sends the prompt tokens and receives the response tokens.

# Authentification (une seule fois)
ollama signin

# Lancer un modèle hébergé
ollama run gpt-oss:120b-cloud

#Which models run in the cloud

The cloud catalog targets precisely what performs poorly locally. It evolves, but the logic remains the same: open-weight or open models, very large models, or multimodal models demanding on VRAM.

100B+ parameter models

Types such as gpt-oss:120b, qwen3-coder:480b, and kimi-k2, which require 60 to 250 GB of VRAM in Q4, are beyond the reach of an individual workstation.

Frontier MoE models

DeepSeek V3/V4, Llama 4 Maverick: the full versions, not the 7B/32B distillations installed locally.

Large specialized models

Coding giants, long-context reasoning (“thinking”) models, high-resolution vision models.

The “normal” models (Mistral 7B, Llama 3.1 8B, Qwen 14B, Phi-4, etc.) are still recommended locally: they fit on a 12 GB RTX 3060 or a MacBook Air, and there is no benefit to offloading them.

#Pricing and limits

Ollama Cloud sits between a flat monthly subscription (with a requests-per-hour quota) and a pay-as-you-go billing model for intensive use. Exact pricing changes over time — check ollama.com/cloud before committing — but a few constants are worth noting.

Hourly quotas

Basic plans limit the number of queries per hour or per day. Comfortable for human chat use, but quickly exhausted by a script looping over 10,000 documents.

Inference-duration billing

Very large models are sometimes billed by GPU time consumed. A long-context request (32k tokens, deep reasoning) mechanically costs more than a “hello.”

No per-token price displayed by default

Unlike OpenAI or Anthropic, the API does not consistently charge per token on the Ollama Cloud side. This makes budgeting less precise.

No consumer-grade SLA

The offering is young, and (at the time of writing) there is no availability guarantee comparable to those of the major clouds.

#Privacy: what really changes

This is where the difference from self-hosting becomes structural, not incidental. When you use Ollama locally, your prompts never leave port 11434 on localhost — this can be verified at the firewall. In the cloud, they travel over the Internet and are processed on third-party infrastructure.

Your prompts leave the perimeter

Any data sent — a contract excerpt, proprietary source code, patient file, or email — leaves your machine. For a law firm, physician, HR department, or R&D lab, that is a deal-breaker.

Retention policy

Ollama states that it does not train its models on your prompts. However, retention for debugging, abuse detection, or logs varies by plan. Read the terms of service before sending sensitive data.

GDPR and hosting

Ollama Cloud servers are primarily in the United States. For European personal data, this raises questions about transfers outside the EU that self-hosting eliminates in one stroke.

No infrastructure-side audit possible

Locally, you can tcpdump port 11434 and prove that nothing leaves the machine. In the cloud, you have to trust the provider—there is no technical way to verify what is logged.

#Self-hosted alternatives for each cloud model

For nearly all models offered in the cloud, there is a local alternative that is either equivalent (a smaller model from the same family) or acceptable (another comparable open-weight model). Here are the useful equivalents.

gpt-oss:120b-cloud → llama3.1:8b or qwen2.5:14b locally

For 90% of conversational use cases, an 8B–14B Q4 model on a RTX 3060 12 GB card does the job. The difference is most noticeable on long reasoning tasks or encyclopedic knowledge.

qwen3-coder:480b-cloud → qwen2.5-coder:14b or 32b

The 32B in Q4 (≈19 GB VRAM) runs on a RTX 4090 or a Mac M-Max. For code autocompletion in an IDE, that's more than enough — see the Continue.dev guide.

deepseek-v3:671b-cloud → deepseek-r1:32b or distilled 70b

Distilled versions of DeepSeek R1 reproduce 80–90% of the full model's reasoning on common benchmarks, locally on a 4090 or Mac Studio.

Huge vision models → qwen2.5-vl:7b or llama3.2-vision:11b

For OCR, image tagging, and description, these two models run on 8–12 GB of VRAM. See the guide to local multimodal vision LLMs.

1M-token context → local 128k models + chunking

Very few real-world use cases require 1M tokens at once. A good RAG pipeline with chunking + reranking can process massive corpora with a local 32k model.

#Decision table

To decide between Ollama Cloud and self-hosting, ask yourself these five questions in order. The first answer of “self-host” settles the debate.

1. Is the data sensitive?

Proprietary code, customer data, healthcare, legal, HR, R&D → self-host, period.

2. Are you subject to the GDPR or industry-specific regulations?

Hospital, bank, public sector, European data → self-host, or dedicated sovereign cloud (not Ollama Cloud).

3. Is the volume predictable and high?

More than a few thousand calls per day → a €1,500 GPU pays for itself within a few months compared with a variable cloud bill.

4. Do you need to operate offline?

Unreliable connectivity, field demos, air-gapped compliance → self-hosting required.

5. Do you absolutely need a 100B+ model?

If so—and only if so—and after testing a local 32B: Ollama Cloud becomes an option worth considering.

#When to use one, and when to use the other

#The few good use cases for Ollama Cloud

One-off evaluation of a giant model

You want to test in 10 minutes whether a 480B coder is worth it before investing in a machine — the cloud avoids an impulse purchase.

Client demo with no on-site hardware

A salesperson showing a POC on their laptop without bringing along a Mac Studio.

A single highly demanding task

A 500-page report to summarize once per quarter, with a long context and no confidential data in it.

Learning and exploration

Discover what a 670B model can do, then return to local models knowing what to aim for.

#When self-hosting wins

Any recurring use

Daily coding assistant, internal team chat, RAG over company documents — predictable costs matter.

Any use involving sensitive data

No debate: legal, healthcare, HR, finance, R&D, and proprietary code stay local.

Any automated use

Batch pipelines, looping agents, n8n, ETL scripts: cloud quotas explode and costs become unpredictable.

Any offline or edge use

Raspberry Pi in the workshop, laptop on the road, site without reliable Internet.

Any need for customization

Fine-tuning, a custom Modelfile, team system prompts, deep integration with your tools—the cloud does not allow it.

#Go further

If you conclude that self-hosting is the right path for your use case (which is likely), these guides will help you get started on the right foot:

Choose the right GPU

The Choosing Your GPU for Local AI guide details the VRAM/budget/performance trade-offs, from the RTX 3060 12 GB to the Mac Studio Ultra.

Understanding quantization

The Choosing Your Quantization guide (Q4, Q5, Q8, FP16) explains how to fit a 32B model on a 16 GB GPU without degrading quality.

Operational privacy

The Privacy Checklist lists the concrete checks needed to ensure that no packet leaves your machine — proof based on facts.