By FreePortal Team · Reviewed by FreePortal Editorial Team · Last updated October 2026

How to Download and Install Ollama Free Guide

Run powerful AI models like Qwen 3.6, Gemma 4, and DeepSeek-R1 directly on your PC — no internet or API keys required.

Is Ollama Free?

Yes, completely. Ollama is free and open source under the MIT licence. There is no paid tier, no subscription, no account requirement for local use, and no restriction on commercial use. The MIT licence permits commercial use, modification, and distribution.

Models that run through Ollama have their own licences — most open-weight models like Qwen, Gemma, and DeepSeek have permissive or research-friendly licences, but some models (notably Meta's Llama models) have usage restrictions for very large-scale commercial deployments. Check the model's licence on its model card before using it commercially at scale.

Current stable release: Ollama 0.40.8 (21 October 2026).

System Requirements

OS: Windows 10/11 (x64 and ARM64), macOS 14 Sonoma or newer, or Linux (Ubuntu 18.04 / 20.04 / 22.04+ or modern 64-bit distributions)
CPU: 64-bit processor with AVX2 support (Intel/AMD). Apple Silicon is fully supported.
RAM: 8 GB minimum for 3B–4B models; 16 GB recommended for 7B–8B and 20B models; 32 GB for 27B–30B models; 64 GB+ for 70B models on CPU.
GPU (optional): NVIDIA GPU with 8+ GB VRAM recommended for 7B–13B models; 16 GB VRAM for 30B models; 32 GB+ for 65B–70B models. NVIDIA driver 550+ required; driver 570+ required for compute capability 5.0–6.2. AMD is supported via ROCm; Apple Silicon uses Metal.
Storage: 10+ GB free space (models range from 2 GB to 40 GB+).

How to Install Ollama

Go to the official site. Visit ollama.com.
Download your version. Click “Download for Windows”, “macOS”, or “Linux” — the site auto-detects your OS. On Windows ARM64 (Snapdragon X Elite), download the native ARM64 build. On Linux, the curl install script is: curl -fsSL https://ollama.com/install.sh | sh. On Windows PowerShell, the equivalent is: irm https://ollama.com/install.ps1 | iex.
Run the installer. On macOS, mount the .dmg and drag Ollama to Applications. On Windows, run the native installer. On Linux, the curl script installs the binary and sets up a systemd service.
Launch Ollama. On Windows and macOS, open the app from your Start menu or Applications folder. On Linux, start the server with ollama serve if it is not already running as a service.

Getting Started with Local Models

Pull a model. Open your terminal and run ollama pull gemma4 or ollama pull qwen3.6:27b — this downloads the model to your machine.
Run the model. Type ollama run gemma4 — you'll see a prompt (>>>) where you can type messages and get responses.
Use the API server (optional). Ollama runs an OpenAI-compatible API at http://localhost:11434/v1 (and a native API at /api) — great for connecting to Python scripts or IDEs.
Try one-shot prompts. Run ollama run gemma4 "Explain how attention mechanisms work in transformers" to get a single answer without entering chat mode.
Manage your models. Use ollama list to see installed models, ollama rm to remove them, and ollama ps to see running models and whether they are on GPU or CPU.

Model Recommendations by RAM

Choosing the right model size matters more than anything else in local AI. These are practical guidelines based on the 2026 Ollama library and typical hardware. Ollama defaults to Q4_K_M quantization.

8 GB RAM — 3B–4B models: llama3.2:3b, qwen2.5:3b, gemma4:e2b. Fast and capable for general chat, but limited reasoning depth. 7B–8B models can run at Q4 if you close other applications, but 16 GB is the comfortable floor.
16 GB RAM — 7B–8B and 20B models: llama3.1:8b, qwen3.6:8b, gpt-oss:20b (MXFP4, 16 GB). The sweet spot for quality vs. speed. GPT-OSS 20B is a notable 2026 addition that fits comfortably in 16 GB.
32 GB RAM — 27B–30B models: qwen3.6:27b (17 GB), qwen3-coder:30b (19 GB), glm-4.7-flash (19 GB), devstral:24b (14 GB). Noticeably better reasoning and coding quality than 7B–8B.
64 GB RAM (or 24 GB+ VRAM) — 70B models at Q4: llama3.3:70b (~43 GB), deepseek-r1:70b (~42 GB). Close to commercial model quality for most tasks. Without a GPU, expect slow CPU inference.
128 GB RAM (or 2× 24 GB VRAM) — 70B models with headroom, and larger MoE models at low quant. The largest open models that fit on consumer hardware. A GPU is strongly recommended at this tier.

Start one size below what you think your hardware can handle — you can always pull a larger model later once you know what works.

Useful Tips

🧠 Start small. If you have 8 GB RAM, begin with 3B–4B models like llama3.2:3b or gemma4:e2b. For 16 GB RAM, try gpt-oss:20b or qwen3.6:8b.
🚀 Enable GPU offload. If you have an NVIDIA GPU, Ollama will use it automatically — but confirm the driver is 550+ (570+ for compute capability 5.0–6.2). Without a supported driver, Ollama silently falls back to CPU.
💾 Download models once. Models are saved locally — you can run them anytime without re-downloading.
🔌 Connect to code editors. Use the OpenAI-compatible endpoint at http://localhost:11434/v1 to integrate with VS Code, Cursor, or Python scripts — no API keys needed. See our Cursor guide.
🔄 Try different models. Use ollama pull to download multiple models and switch between them easily.
🔒 Private and offline (for local models). Local inference stays on your machine. If you sign in to Ollama Cloud, cloud-model prompts are sent to Ollama's servers.
⚙️ Control context size. During a session, use /set parameter num_ctx 8192. To set a default, use the OLLAMA_CONTEXT_LENGTH environment variable or add PARAMETER num_ctx 8192 to a Modelfile. Larger context uses more RAM.
📊 Monitor system resources. Run ollama ps in another terminal to see which models are loaded and whether they are on GPU or CPU.
🖥️ Check your NVIDIA driver. Run nvidia-smi. If the driver is below 550 (or below 570 on older cards), update it before troubleshooting GPU performance.

Ollama vs LM Studio vs GPT4All

Ollama: MIT-licensed, open-source Go daemon, free for personal and commercial use. Simple CLI (ollama pull, ollama run), excellent developer integration (Cursor, Continue, LangChain), Docker support, and an OpenAI-compatible API on localhost:11434/v1. Best for developers and anyone who wants to automate local AI from scripts, containers, or IDEs.
LM Studio: Closed-source desktop GUI, also free for personal and internal business use (since 8 July 2025). Built-in model discovery, visual parameter controls, and a toggle-on OpenAI-compatible server on port 1234. Faster out-of-the-box Apple Silicon inference via its MLX backend. Best for users exploring local AI with a visual interface — see our LM Studio guide.
GPT4All: Free, open source, and lighter than both. Bundled with a local document chat feature (RAG) and model discovery. Emphasises CPU inference and is the simplest “install and chat” option. Best for users who want a basic local chat app without much configuration.

If you want to build with local models from scripts, Docker, or IDEs, Ollama is the developer standard and is fully open source. LM Studio is best for GUI exploration and is also free for work. GPT4All is the lightest and simplest option for pure chat.

Troubleshooting Common Issues

❌ “ollama: command not found” after install. The install script adds Ollama to your PATH, but you need a new terminal session (or a shell restart) to pick it up. Close and reopen your terminal, or run source ~/.bashrc (or equivalent for your shell).
🧠 Model runs out of memory during load. The model is too large for your RAM+VRAM. Try a smaller model, a lower quantization tag (e.g., :q4_K_M), or reduce the number of GPU layers by creating a Modelfile with PARAMETER num_gpu 20 (lower than the default) so more of the model stays on CPU (slower but works).
🐌 Responses are extremely slow. The model is running on CPU. Run ollama ps to check the PROCESSOR column. If it shows CPU, confirm your NVIDIA driver meets the minimum (550+, or 570+ for compute capability 5.0–6.2). On Apple Silicon, unified memory handles this automatically.
💾 Downloads stall or fail. Model downloads come from Ollama's registry (not Hugging Face by default). If downloads consistently fail, check your network — corporate firewalls and VPNs sometimes block them. Try ollama pull again with a different model to see if it is model-specific.
🔌 API server not responding. Check that the server is running with ollama serve (or that the app is running on Windows/macOS). Test with curl http://localhost:11434/api/tags in a terminal — a response means the server is working. For OpenAI-compatible clients, use the base URL http://localhost:11434/v1.
🎨 Output is cut off mid-sentence. Usually the context window is exhausted. Start a new chat, or increase context length with /set parameter num_ctx 8192 during the session (or set OLLAMA_CONTEXT_LENGTH before starting the server).
🖥️ GPU not detected. Update your NVIDIA driver to 550+ (570+ for older cards). Confirm the GPU has compute capability 5.0+. On Windows ARM64, use the native ARM64 build to get CUDA support.

Frequently Asked Questions

Can I run Ollama without a GPU?

Yes, but it will be slower. Ollama falls back to CPU inference (AVX2 required), which works but produces responses in tens of seconds rather than seconds. For occasional use, this is fine; for regular use, an NVIDIA GPU (8+ GB VRAM, driver 550+; driver 570+ for compute capability 5.0–6.2) or Apple Silicon Mac is much faster.

What is the difference between Ollama and LM Studio?

Ollama is a MIT-licensed, open-source Go daemon built for scripts, Docker, and API integration. LM Studio is a closed-source desktop GUI built for model browsing and interactive chat. Both are free for personal and internal business use — LM Studio removed its separate commercial licence on 8 July 2025. Choose Ollama for automation and open-source licensing; choose LM Studio for a visual interface and faster out-of-the-box Apple Silicon inference. See our LM Studio guide.

How big are the model downloads?

It depends on model size and quantization. Small 3B–4B models are around 2–3 GB. 7B–8B models are around 4–5 GB. 13B–14B models are 8–10 GB. 27B–30B models are 17–19 GB at Q4. 70B models at Q4 are 40–43 GB. Plan storage and download time accordingly.

Does Ollama send my data anywhere?

For local models: no. Prompts, responses, and inference happen on your machine. The only network traffic is when you download models from Ollama’s registry. If you sign in to Ollama Cloud, prompts for cloud models are sent to Ollama’s servers — local models remain private.

Can I use local models with Cursor or VS Code?

Yes. Ollama exposes an OpenAI-compatible API at http://localhost:11434/v1 (and a native API at /api), which Cursor and the Continue extension in VS Code support. Configure your editor to use the local Ollama endpoint — no API keys, no cost, and your code never leaves your machine.

Does Ollama support vision models?

Yes. Current multimodal options include Gemma 4 (vision and audio) and Qwen 3.8:27b (text and image input). Run ollama run gemma4 and provide an image path when prompted, or use the API with base64-encoded images. Older options like llava and llama3.2-vision still work but are no longer the best defaults.

Does Ollama require an account?

No account is required for local models. Ollama Cloud (beta) requires sign-in and an API key, but the local daemon and CLI work without any account.

← Back to all resources