Run Ollama Locally: Hardware Requirements in 2026 (RAM, GPU, CPU)

The real hardware requirements for Ollama in 2026, by model size, with VRAM and RAM numbers.

## Run Ollama Locally: Hardware Requirements in 2026 Before you pull a model, check whether your machine can actually run it. These are the practical requirements in 2026. ### The golden formula Model size (GB) = parameters (B) x bytes per weight. At 4-bit quantization: 7B model = about 4-5 GB, plus 1-2 GB context overhead. ### Requirements by model size | Model | Quantized size | Minimum RAM/VRAM | Experience | |-------|---------------|-----------------|------------| | 0.5-1B | ~0.5-1 GB | 4 GB | instant, low quality | | 3B | ~2 GB | 6-8 GB | fast, decent | | 7B | ~4-5 GB | 8 GB VRAM or 16 GB RAM | the sweet spot | | 13B | ~8 GB | 16 GB VRAM | good quality | | 32B | ~19 GB | 24 GB VRAM or 32 GB RAM | high quality, slow CPU | | 70B | ~40 GB | 48 GB+ | flagship, needs serious hardware | ### CPU vs GPU vs Apple Silicon - **GPU (CUDA)**: fastest. RTX 4060 (8 GB) runs 7B Q4 at 30-50 t/s - **CPU only**: 7B Q4 on a modern 8-core desktop gives 4-8 t/s - usable for chat, slow for long docs - **Apple Silicon**: 16 GB Mac runs 13B Q4 at 15-25 t/s thanks to unified memory - **RAM matters more than CPU cores**: for CPU inference, fast dual-channel DDR5 helps most ### Check before you install [bash] # Linux: check GPU nvidia-smi # macOS: check unified memory sysctl hw.memsize # free RAM free -h [/bash] ### FAQ **Can I run a 7B on 8 GB total RAM?** Tight - use 4-bit and close other apps; 16 GB is the comfortable minimum. **Does context length matter?** Yes - each token of context uses ~4 bytes of memory in most runtimes; a 128k context adds significant RAM.