Frequently asked questions

The context window is the maximum amount of text — measured in tokens — that a model can take into account at once. It covers your input, the conversation so far, and the model's generated answer together. A model with a 1M context can process roughly several hundred pages in a single request; a model with 128K handles far less. Long documents, large codebases, and multi-step conversations need a larger window. If a request exceeds the window, the oldest content must be trimmed, and the model loses sight of it.

Modalities describe the kinds of data a model can work with. Text means the model reads and writes text only, matching Ollama's default. Text, Image means the model is multimodal: you can include images in your prompt, and the model reasons over them alongside textual instructions. All Ollama cloud models output text even when they accept image input.

Size is the total number of parameters the model was trained with, which is the standard measure of its capacity — for example 20B means 20 billion parameters. Larger models are generally more capable but slower and more expensive to run. Some entries show an active-parameter count too, such as 120B (12B active): those models use a mixture-of-experts (MoE) architecture where only part of the network runs per token, giving big-model quality at reduced cost. Extremely large models are labeled in trillions, e.g. 1.65T.

Usage level is the badge with the four-bar meter shown next to each model name. It indicates the demand tier the model sits in on Ollama's cloud infrastructure, and it is also the filter used on the registry's front page. The levels are: Low (green, one bar) — light demand, usually fast and cheapest to run; Medium (blue, two bars) — moderate demand; High (amber, three bars) — heavy demand, expect more queuing under load; Extra high (orange, four bars) — the most sought-after models, where capacity pressure is greatest. Token-based (purple, dashed meter) is different: instead of a demand tier it means the model is billed per token on Ollama's cloud, so a monthly price label shows in the model list. Levels are sourced from the model's ollama.com library page and refreshed hourly, and can be overridden locally for models you run yourself.

Pulls counts how many times the model has been downloaded or invoked from ollama.com, refreshed hourly from the model's live library page. It is a rough popularity metric: a model with 41.0M pulls is used far more than one with 698.5K. High pull counts usually indicate a mature, well-tested model, but they don't say anything about fit for your specific task.

Tags are the published variants of a model on ollama.com — different quantizations, sizes, or revisions of the same family, such as latest, a smaller quant, or a dated snapshot. A higher tag count (28 vs 2) means more choices when pinning a specific version on Ollama's hosted infrastructure. This registry always tracks the :cloud tag, which runs the default variant on Ollama's servers.

Updated shows how long ago the model's page on ollama.com was last changed — typically a re-release, a new tag, revised model card content, or refreshed benchmarks. Updated 3 d ago means the publisher touched the model page three days ago. A recently updated model is actively maintained; a model updated many months ago may still work fine but is unlikely to receive new capabilities. The value is scraped hourly together with pulls and tags.