What models can I run with Ollama?

Ollama supports a curated library of open-weights models including Llama, Qwen, Gemma, DeepSeek, Mistral, and Phi. It handles quantized GGUF builds with sensible defaults for easy execution.

How does Ollama integrate with existing development tools?

It exposes a local OpenAI-compatible HTTP API, allowing clients like Open WebUI, Continue, LangChain, and custom apps to connect without code changes. The same API works for both local and cloud tiers.

Is Ollama free to use?

Local use is free and open source under the MIT license. Paid cloud tiers are available for Pro at $20/mo and Max at $100/mo, which allow running larger models concurrently in US, EU, or Singapore regions.

What are the limitations of using Ollama?

It uses llama.cpp under the hood, which may be slower than vLLM or SGLang for high-throughput production serving. The cloud tier is newer and less battle-tested than some competitors, and GPU offload configuration can be finicky.

Ollama

The de facto runtime for running open-weights LLMs locally, now with a paid cloud tier for bigger models.

Freemium· Free local; Pro $20/mo; Max $100/moCodingMulti-model (Llama, Qwen, Gemma, DeepSeek, Mistral, Phi, etc.)7.2 / 10

Visit website →

In short

Ollama simplifies downloading and serving open-weights LLMs locally on Mac, Linux, and Windows. It exposes an OpenAI-compatible API for seamless integration with existing tools. A paid cloud tier is available for larger models that exceed local hardware limits.

Best for

Pick Ollama if you want one command to run open-weights LLMs on your own hardware with a clean API that any client can hit.

Skip if

Skip it if you need high-throughput production inference at scale or specifically want frontier closed models like GPT-5 or Claude.

Ollama is the simplest way to download, run, and serve open-weights large language models on your own machine. A single CLI pulls quantized GGUF builds of Llama, Qwen, Gemma, DeepSeek, Mistral, Phi and dozens of others, then exposes them over a local OpenAI-compatible HTTP API so any client (Open WebUI, Continue, LangChain, Claude Code, custom apps) can talk to them without code changes. Mac, Linux and Windows are all first-class.

The tool is open source (MIT) and free for local use, which is how most people meet it. Ollama now also sells a hosted tier for models too big for a laptop: Pro at $20/mo runs up to 3 cloud models concurrently, Max at $100/mo bumps that to 10, with US/EU/Singapore regions and a 'your data is never trained on' promise. The hybrid story (develop locally, burst to cloud with the same API) is the differentiator versus pure-local llama.cpp or pure-cloud Together/Fireworks.

The ecosystem is huge: every desktop LLM client, IDE plugin, and agent framework worth mentioning has an Ollama adapter, and the model library is curated and quantization-tagged so 'ollama run qwen2.5:14b' Just Works. Limitations: it's a wrapper over llama.cpp under the hood, so cutting-edge inference engines (vLLM, SGLang) are faster for serving, and the cloud tier is newer and less battle-tested than Together/Groq.

Editor's take

Ollama is the package manager for local LLMs and has basically won that category. The new paid cloud tier is a smart hedge for users whose laptops can't fit a 70B model, though serious production teams will still reach for vLLM or a hosted inference API. For everyone else, it's the obvious default.

— The AI Tool Bible editorial team

Pros

✅ Easiest path to running open-weights LLMs locally on Mac/Linux/Windows
✅ OpenAI-compatible API means existing tooling works out of the box
✅ Huge curated model library with sensible quantization defaults
✅ Same API for local and cloud lets you scale without rewriting code
✅ Open source (MIT) with a massive integration ecosystem

Cons

⚠️ Underlying llama.cpp engine is slower than vLLM/SGLang for production serving
⚠️ Cloud tier is newer than competitors like Together or Fireworks
⚠️ Configuration of GPU offload and context length can be finicky

Use cases

local-llmself-hosted-inferenceprivate-coding-assistantrag-backendoffline-ai

Frequently asked

What models can I run with Ollama?: Ollama supports a curated library of open-weights models including Llama, Qwen, Gemma, DeepSeek, Mistral, and Phi. It handles quantized GGUF builds with sensible defaults for easy execution.
How does Ollama integrate with existing development tools?: It exposes a local OpenAI-compatible HTTP API, allowing clients like Open WebUI, Continue, LangChain, and custom apps to connect without code changes. The same API works for both local and cloud tiers.
Is Ollama free to use?: Local use is free and open source under the MIT license. Paid cloud tiers are available for Pro at $20/mo and Max at $100/mo, which allow running larger models concurrently in US, EU, or Singapore regions.
What are the limitations of using Ollama?: It uses llama.cpp under the hood, which may be slower than vLLM or SGLang for high-throughput production serving. The cloud tier is newer and less battle-tested than some competitors, and GPU offload configuration can be finicky.