Ollama runs open models locally: no cloud, no API bill, nothing leaves your machine. The hard part is picking the right one for your RAM, not the biggest one you've heard of. Here's the honest shortlist as of July 2026, what each is actually good at, the real download sizes, and the two commands to get serving.
Ollama is a one-command runner for open LLMs. Install it, run a model, and it downloads the weights and serves a local API. No cloud, no API key, nothing leaves your machine.
Every model is a download that has to fit in memory. The file size is the floor, and you want headroom on top for context and the OS. That is why the picker below is organised by RAM, not by benchmark.
Once a model is running it speaks an OpenAI-compatible API at localhost:11434, so most AI apps and scripts can point at it with a one-line base-URL swap.
Local models are genuinely useful and private, but a 12 to 32B model on your laptop is not GPT-5 or Claude Opus. Match the model to the task and you will be happy. Expect frontier quality and you will not.
Sizes are the actual Ollama downloads as of 16 Jul 2026. The RAM tier is the file plus headroom for context and the OS, so treat it as a floor, not a promise. On 16 GB, gemma4:12b is the comfortable pick and gpt-oss:20b works if you close memory-hungry apps.
Download it from ollama.com for macOS, Windows, or Linux. It installs as a background service that is ready the moment it finishes.
ollama run gemma4 downloads the default pick and drops you straight into a chat. The first run pulls the weights, and after that it loads from disk instantly.
Ollama is now serving an OpenAI-compatible API at localhost:11434. Set that as the base URL in any app or script and you are running local.
# 1 · download Ollama from ollama.com (macOS / Windows / Linux) ollama run gemma4 # pull + chat with the default pick # 2 · it is now serving at localhost:11434 (OpenAI-compatible) curl localhost:11434/api/tags # list the models you have pulled
A 7.6 GB model needs more than 7.6 GB of RAM once you add context and the OS. On 16 GB, stick to the smaller picks and close the memory-hungry apps.
These are excellent 12 to 32B models, not GPT-5 or Claude Opus. Great for drafts, code help, private data, and offline work. Weaker on long, hard reasoning chains.
The first ollama run downloads gigabytes. After that the model loads from disk and runs at your hardware speed, which on recent Apple Silicon is genuinely usable.
Ollama ships quantized weights so models fit consumer RAM. That is the right trade for local use, but it is a small quality cut versus the full-precision cloud version.
The picker above is a snapshot. The full Ollama library lists every model, all its size variants, context windows, and whether it sees images, updated as new models land. Check the exact download before you pull.