Run ollama list to see the installed model tags. Identify the exact tag your projects no longer use, remove it with ollama rm MODEL:TAG, and check the list again. Compare free space before and after. Models can share stored data, so the size displayed for one tag may differ from the amount freed.
Make an inventory before removing anything
Run ollama list on the computer that stores the models. Record the exact names and tags displayed. Similar names may represent different parameter sizes, quantizations, or versions. A script that specifies a particular tag can break if you remove that exact item while keeping another with a familiar name. If several people use the workstation, ask which models are still in active use before treating an old download date as evidence that it is disposable.
Check which Ollama instance you are inspecting. A client can point at a different host from the machine where you are typing the command. If your app uses the local endpoint, confirm the interface and the model tag with a small request. Our local Ollama Chat API guide shows how to send a known request and inspect the returned model. Record the model and environment for each workflow you intend to keep. This takes less time than trying to reconstruct the names after deleting several tags together.
Check which files and settings are replaceable
A published model you can download again is different from a model you built from a private file, a custom adapter, or a specific Modelfile. Keep the original inputs and configuration needed to recreate custom models. Check that you can find those inputs somewhere other than the installed model store. For a published tag, verify availability if you rely on reproducing an older benchmark. A later download can take time, and the tag you expect may not be available forever.
Review your shell scripts, app settings and any benchmark notes for references to the tag. Examples and test fixtures can be easy to overlook if they are run infrequently. For model comparisons, keep a record of which quantization belonged to each result. The Ollama quantization guide explains why a similar tag can behave differently. If you have two variants used for different workloads, test the remaining one before removing the other. This gives you a clear way to reverse the decision if a real workload needs it.
Measure free space on the right drive
Use your operating system’s disk view to record free space on the volume storing the models. If the storage location was changed, check that you are measuring the destination volume rather than the system drive by habit. Avoid starting a big download or cleanup while you are measuring. A simultaneous update can obscure a small space gain. Write down the free space, then check again under comparable conditions after removing a model.
The model list helps you identify large candidates. Treat its size as a guide to the data associated with that tag. It is not a promise that removing the tag will free exactly the displayed amount. Stored layers can be shared, so data another installed tag still needs can remain. Compare the real free space change and the remaining model list rather than expecting the arithmetic to balance for every removal. If space is still tight, review other usage on that volume using the operating system’s own tools.
Remove one exact tag and verify it
Finish requests using the model before removing it. If it is currently loaded, use ollama ps to identify it and the supported stop command if you need to unload it. Unloading affects the running model, while removing addresses the installed copy. Our Ollama processor diagnostics guide explains how to inspect loaded models. Then run ollama rm MODEL:TAG with the exact name you intend to remove. Replace the example with the value from your own list. If you are automating the same operation, refer to the official delete endpoint documentation and handle errors before proceeding to another tag.
Run the list command again and confirm the removed tag is absent and the expected ones remain. Test an ordinary request against a model you kept. If a workflow now fails, look for an explicit reference to the removed tag before changing the server configuration. Pull the appropriate published tag again if you need it, or rebuild from your retained custom inputs. A single removal and a known request are enough to check the process before doing more.
Understand what remains after cleanup
If little disk space returns, check for a related tag that uses the same model data. Verify that the removal succeeded and that the measured drive was the right one. Look at operating system activity or another large download that could have used space during the test. Avoid manually deleting shared model files from the storage directory. The supported model controls understand the installed references, while a manual removal can break tags you intended to keep.
Disk storage and memory are separate. Removing an unused tag can reduce disk use. Unloading a running model can release resources even if you keep the installed files. Choose the operation based on which resource is constrained. If your storage view shows that the system drive is full for other reasons, use a targeted cleanup such as the process in our Windows storage guide instead of removing useful models indiscriminately.
Make the next cleanup easier
Keep a short inventory of the models you use for chat, embeddings and benchmark comparisons. Put the tag beside the project that depends on it. Before downloading several variants, estimate the free space needed and keep room for normal application updates and temporary files. After experimenting, remove one candidate and record the actual space gained so you can decide whether further cleanup is worth doing.
Consider checking changes after a major upgrade or a migration between computers. Run one representative request for each important workflow and confirm the expected model is installed. If the computer must run offline later, test it without a network before removing anything else. Restoring a downloaded model is straightforward only when you can reach the source again.
If you are working with a script that periodically pulls models, review that script after cleanup. It may immediately download a tag again if the configuration still names it. Update the project’s intended model explicitly and perform a normal test run before judging whether the cleanup held.
Compare free space again after the machine settles. If a background task was still writing files, wait for it to finish and record the new reading before deciding whether another model needs to go.
Frequently asked questions
Does stopping a model delete it?
Stopping unloads the running model. Use the supported remove command to remove an installed tag.
Will the full displayed size be freed?
Check the actual free space change. Related installed tags can share data.
Can I download the model again?
A published tag may be downloadable again if the source still offers it. Keep the inputs required to rebuild custom models separately.
