Ollama is an open-source platform for running large language models (LLMs) locally on macOS, Linux, and Windows. It lets you download, manage, customize, and run models such as Llama 4, Qwen 3, Gemma 3, Phi-4, DeepSeek-R1, Mistral, and many others, without relying on cloud services. Ollama is free and open-source under the MIT License.

Do I need an internet connection to use Ollama?

Only to download models. Once installed, these models can operate entirely offline without sending prompts or responses to cloud servers. However, downloading models, accessing cloud-hosted models, and using online integrations require an internet connection.

Can I use Ollama with my own models?

Yes, Ollama supports custom models via GGUF format and lets you tweak prompts and parameters through a Modelfile.

Features

  • Local-first and lightweight: Install and run LLMs via a simple CLI (ollama run ), leveraging your own CPU/GPU instead of cloud servers.
  • Extensible and developer-friendly: Integrates well with Docker, REST API, and libraries in Python and JavaScript.
  • Customizable models and prompts: Supports tweaking parameters, system messages (via "Modelfile"), and importing custom GGUF/Safetensors models.

What's New

  • Added support for Ripple Preview, a 65B model designed for agentic coding.
  • Reduced memory usage for Qwen3.8 Flash Next by approximately 27 GiB on Apple Silicon.
  • Improved Claude Code integration through ollama launch claude.
  • Added clearer messages when cloud models have been retired.
  • Improved compatibility with the OpenAI Responses API.
  • Added CUDA 13.5 support for Windows on ARM.
  • Updated llama.cpp.

System Requirements

Platform & OS

  • macOS: Big Sur (11) or later
  • Linux: Ubuntu 18.04 / 20.04 / 22.04 or later
  • Windows: Full support via native installer or WSL 2

CPU

  • Minimum: Any modern x86-64 CPU supporting AVX2 (Intel/AMD)
  • Recommended: 11th-gen Intel or AMD Zen 4 with AVX512 and ideally DDR5 memory

Memory RAM

  • Minimum: 8GB
  • Recommended: 16GB

GPU (Optional, but boosts performance significantly)

  • Not required, but speeds inference dramatically
  • Minimum recommendation: Nvidia or AMD GPU with 8 GB VRAM for 7B – 13B models

For large models:

  • 30B model: 16 GB VRAM
  • 65B model: 32 GB VRAM
  • Very large models (70B+): 48 GB or more, possibly multi-GPU