OllamaNew

Updated

An open-source tool for running large models locally; pull and execute Llama, Qwen, and others with a single command, or call its hosted cloud models per token.

  • Popularity
  • Coding
  • Free tier
Visit website
Ollama interface preview
Report incorrect information

We send only this page address and the issue type to the editorial review queue. No account or contact details are needed.

At a glance

  • Free tierPartialLocal runs are free and unlimited; cloud models are billed against credits
  • Open sourceYes
Pricing

Ollama itself is open source and free, and running models on your own hardware is unlimited — the only cost is local compute. Cloud models are billed against monthly credits: Free includes starter credits and a subset of models, Pro is $20/month ($200/year) with $60 of credits, Max is $100/month with $300, Team is $500/month with $1,000 shared, and Enterprise is custom.

Pricing changes over time; check the official site

Comparisons
Tutorials

Running large models on your own computer used to require a mess of environment setup, quantization, and dependency management. Ollama simplifies this into a single command: ollama run qwen3, which automatically downloads the model, loads it, and starts chatting. It is currently one of the most popular tools for running open-source large models locally.

What Is Ollama?

Ollama is an open-source tool for running large models locally, supporting macOS, Windows, and Linux. It accomplishes three key tasks:

  • One-click model pull: Directly download quantized versions of mainstream open-source models like Llama, Qwen, Gemma, DeepSeek, Mistral, and others from the model library.
  • Local inference: Models run on your own machine; data is never uploaded, ensuring full offline availability.
  • Provides a local API: Once started, it exposes a local HTTP interface compatible with OpenAI-style calls, making it easy to integrate local models into various applications and frameworks.

Why Use It

  • Privacy: Sensitive data, code, and documents are processed locally without being sent to any cloud service.
  • Zero Cost: No API call fees; it only consumes your own computing power.
  • Offline Capable: Works even when disconnected from the internet, making it ideal for intranets or secure environments.
  • Customizable: Customize system prompts and parameters via a Modelfile, or even import GGUF weights you have fine-tuned yourself.

It Now Runs Cloud Models Too

Ollama is no longer local-only. It offers hosted models running in the US, Europe, and Singapore, used exactly like local ones — point at the cloud model name, with no code rewrite or new SDK — which helps when a model is too large for your own VRAM. Ollama states that cloud prompts and responses are never logged or trained on and that it requires zero-data-retention policies from its hosting partners. Running models on your own hardware stays free and unlimited; cloud usage draws on paid credits.

Two product changes stood out in 2026: an MLX backend on Apple Silicon that uses unified memory and markedly speeds up decoding on Macs (the preview requires 32GB or more of unified memory), and a Claude Desktop gateway that lets the desktop app talk to your local models.

Who Should Use It

  • Individuals and teams that prioritize privacy and do not want to hand over data to the cloud.
  • Developers building local prototypes, testing different open-source models, or deploying offline.
  • Scenarios requiring low-cost, high-volume model invocation (such as batch processing).

Usage Recommendations

The performance of local models depends on your hardware: more VRAM or system memory allows you to run larger, smarter models. Consumer-grade devices are typically suitable for running medium-to-small-sized models (in the 7B–14B parameter range), while achieving quality close to top-tier closed-source models requires a stronger GPU. Ollama is commonly paired with Open WebUI, LangChain, and various desktop clients to serve as their local inference backend.

Pricing

Ollama itself is open source and free, and running models on your own machine is unlimited — the only cost is local compute. Cloud models are billed separately: since August 31, 2026 they use transparent per-token rates, each plan carries a monthly credit pool, and once it runs out you continue at the same per-token rate. There are no service fees, the earlier 5-hour and weekly usage windows are gone, and unused credits do not roll over. The official pricing page (checked 2026-09-06) lists Free (starter credits plus a subset of models, with all models unlocked once you add credits), Pro at $20/month or $200/year ($60 of credits, multiple concurrent models), Max at $100/month ($300 of credits, 10 concurrent requests), Team at $500/month ($1,000 shared credits, unlimited users), and custom Enterprise pricing. Check the official page for current figures.