Ollama minimum requirements 2026: hardware needed to run local LLMs

Ollama is the easiest way to run local LLMs in 2026. A single command downloads and installs the model, a REST API available immediately, and decent multi-platform support. But "install Ollama" doesn't mean "run Llama 70B without suffering". Hardware still matters enormously. This guide tells you exactly what you need based on the models you'll use.

What is Ollama

Ollama is an open-source LLM runtime built on llama.cpp. It handles download, quantization, context and serving in a single binary. Supports Mac, Windows, Linux and, in 2026, also Snapdragon X2 via native ARM build.

Its advantage over raw llama.cpp is abstraction: you don't touch GGUF files directly, don't compile with CUDA flags, don't fight dependencies. ollama run llama3.1 and you have a model responding.

Minimum requirements by OS

Windows

• OS: Windows 10 22H2 or Windows 11.
• RAM: 8 GB minimum, 16 GB recommended.
• GPU: optional. NVIDIA with CUDA 12+ accelerates massively. AMD with ROCm experimental.
• Disk space: 5 GB for Ollama + model sizes (2-50 GB each).

macOS

• OS: macOS 12 (Monterey) or higher. 14+ recommended.
• Apple Silicon: M1 or higher. M5 ideal.
• RAM (unified): 8 GB minimum, 16 GB recommended, 32 GB+ for 13B-30B models.
• Disk space: 5 GB + models.

Linux

• Kernel: 5.10 or higher.
• RAM: 8 GB minimum, 16 GB recommended.
• GPU: NVIDIA recommended with official drivers and CUDA 12+.
• Disk space: 5 GB + models.

Hardware by model in Ollama

Small models — Llama 3.2 1B, 3B, Phi-3 mini, Qwen 2.5 1.5B

• Minimum RAM: 8 GB.
• VRAM (optional): 2-3 GB for acceleration.
• Throughput without GPU: 15-30 tokens/s on modern CPU.

These models work for autocomplete, classification, simple translation, tag generation. Quality below ChatGPT-3.5 but fast and cheap to run.

Medium models — Llama 3.1 8B, Mistral 7B, Qwen 2.5 7B

• Minimum RAM: 16 GB.
• Ideal VRAM: 6 GB (Q4_K_M) or more.
• Throughput with RTX 4060: 25-40 tokens/s.
• CPU-only throughput (Intel Core Ultra 7): 4-7 tokens/s.

Sweet spot for daily laptop use. Quality close to ChatGPT-3.5 on common tasks.

Large models — Llama 3.1 13B, Qwen 2.5 14B, DeepSeek R1 14B

• Minimum RAM: 32 GB.
• Ideal VRAM: 8-10 GB (Q4) or 11-12 GB (Q5).
• Throughput with RTX 4080: 25-35 tokens/s.

Notable quality, approaching GPT-4 mini in many tasks. Detail in our how much RAM for ChatGPT local guide.

Extra large — Qwen 32B, DeepSeek R1 32B, Llama 3.3 30B

• Minimum RAM: 48 GB.
• Ideal VRAM: 18-20 GB.
• Throughput with RTX 5080: 10-15 tokens/s.

On laptop only viable with RTX 5080+ or Mac M5 Pro/Max.

Massive — Llama 3.3 70B, DeepSeek R1 70B

• Minimum RAM: 64 GB.
• Ideal VRAM: 40+ GB (offload needed on consumer laptop).
• On Mac M5 Max 96 GB unified: 8-15 tokens/s.
• On PC with RTX 5090 (24 GB) + 64 GB RAM: 2-6 tokens/s.

Serious quality. Replaces GPT-4 on many tasks. Limitation: slow except on very powerful hardware.

GPU vs CPU only setup

NVIDIA GPU with CUDA

Ollama automatically detects the GPU and uses it. For models that fit in VRAM, throughput is 5-10x higher than CPU only.

To verify Ollama is using the GPU:

``bash ollama run llama3.1 ``

While it generates, open another terminal:

``bash nvidia-smi ``

You should see ollama using the GPU at 80-100%.

CPU only

Works but limited. For models up to 7B in Q4 it's usable (4-10 tokens/s). For 13B+ without GPU, the experience is unviable except for batch processing.

Apple Silicon

Ollama runs native on Metal (Apple's graphics API). Leverages unified memory: what matters is total RAM, no VRAM/RAM distinction. Works excellently and is the simplest setup without a dedicated GPU.

Useful optimizations

Quantization

Ollama uses Q4_K_M by default on most models. For more quality: ollama pull llama3.1:8b-instruct-q5_K_M. For less memory: q3_K_M (not recommended, quality drops).

Context size

By default Ollama uses the model's context (4K-128K depending on model). Reducing it saves RAM:

``bash OLLAMA_NUM_CTX=4096 ollama run llama3.1 ``

Keep the model loaded

By default Ollama unloads the model from RAM after 5 minutes of inactivity. For interactive use, set keepalive:

``bash OLLAMA_KEEP_ALIVE=24h ``

Comparison with LM Studio and llama.cpp

• Ollama: easiest. Good for developers wanting immediate REST API. Limited in advanced customization.
• LM Studio: nice GUI, ideal for non-CLI users. More visual, less automatable.
• llama.cpp: most flexible. Total control, but requires compilation and complex CLI. For serious production or research.
• vLLM: high performance for serving. For production deployments, not individual use.

For 80% of cases, Ollama is what we recommend.

Practical recommendations by budget

• Try Ollama without spending: any laptop with 16 GB of RAM. Models up to 8B in Q4. Acceptable.
• Comfortable daily use: laptop with RTX 4060 + 32 GB. Fluid 8B-13B models. See what laptop for local AI.
• Professional: laptop with RTX 5080 + 64 GB or MacBook Pro M5 Max + 48-96 GB. Viable 30B-70B models.

FAQ

Does Ollama work without GPU? Yes, but large models are very slow. For 7B without GPU, 5-10 tokens/s. For 13B, practically unviable interactively.

How much disk space do I need? Each model weighs 1-50 GB quantized. If you experiment with several, you'll fill 100-200 GB quickly. Minimum 256 GB free recommended, ideally 1 TB.

Does Ollama send data to the cloud? No. Everything processes locally. One of the main attractions for use with sensitive data.

Can I use Ollama with Open WebUI or similar? Yes. Ollama exposes an OpenAI-compatible API. Open WebUI, LibreChat, AnythingLLM all work with it.

How do I update models in Ollama? ollama pull downloads the latest version. Weights are overwritten.

Which specific laptop serves you for Ollama?

Tell the AI advisor what models you'll run and your budget. You'll get the specific laptop with the VRAM and RAM you need, and expected throughput.