[{"data":1,"prerenderedAt":55},["ShallowReactive",2],{"blog-en-local-ai-models-disk-space":3},{"meta":4,"html":14,"headings":15,"raw":54},{"title":5,"description":6,"topic":7,"date":8,"author":9,"icons":10},"Local AI Models Filling Your Disk? Where Ollama, Hugging Face, and LM Studio Keep Models and How to Remove Them","Find the model folders of Ollama, Hugging Face, and LM Studio, delete models you no longer use with each tool's own commands, and avoid shared-file surprises.","developer-cleanup","2026-10-29","WashBear Team",[11,12,13],"ai-model","model-cube","hard-drive","\u003Cp>A 7B or 8B parameter model still takes 4 to 5 GB after 4-bit quantization. Try a few sizes and quantization levels and tens of gigabytes disappear fast. To make it harder, every tool has its own model folder, and some of them are hidden. This guide lists the default locations of the common tools and how to delete models in each, then explains what WashBear can do for you.\u003C\u002Fp>\n\u003Ch2 id=\"where-models-live\">\u003Ca class=\"heading-anchor\" href=\"#where-models-live\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Where models live\u003C\u002Fh2>\n\u003Cdiv class=\"table-wrap\">\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>Tool\u003C\u002Fth>\u003Cth>Default location\u003C\u002Fth>\u003Cth>Layout\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>Ollama\u003C\u002Ftd>\u003Ctd>\u003Ccode>~\u002F.ollama\u002Fmodels\u003C\u002Fcode>\u003C\u002Ftd>\u003Ctd>\u003Ccode>blobs\u003C\u002Fcode> holds model data; \u003Ccode>manifests\u003C\u002Fcode> records which data makes up each model\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>Hugging Face\u003C\u002Ftd>\u003Ctd>\u003Ccode>~\u002F.cache\u002Fhuggingface\u002Fhub\u003C\u002Fcode>\u003C\u002Ftd>\u003Ctd>One folder per repository; models start with \u003Ccode>models--\u003C\u002Fcode>, datasets with \u003Ccode>datasets--\u003C\u002Fcode>\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>LM Studio\u003C\u002Ftd>\u003Ctd>\u003Ccode>~\u002F.lmstudio\u002Fmodels\u003C\u002Fcode>\u003C\u002Ftd>\u003Ctd>Two levels: publisher, then model\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003C\u002Fdiv>\n\u003Cp>Python libraries that download models through Hugging Face, such as transformers and diffusers, share the same cache. All of these locations can be changed: Ollama uses the \u003Ccode>OLLAMA_MODELS\u003C\u002Fcode> environment variable, Hugging Face uses \u003Ccode>HF_HOME\u003C\u002Fcode> or \u003Ccode>HF_HUB_CACHE\u003C\u002Fcode>, and LM Studio lets you change it on the app's \u003Cstrong>My Models\u003C\u002Fstrong> page.\u003C\u002Fp>\n\u003Ch2 id=\"before-you-delete\">\u003Ca class=\"heading-anchor\" href=\"#before-you-delete\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Think before you delete\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Cstrong>The cost of downloading again\u003C\u002Fstrong>: models are often several gigabytes, and re-downloading on a slow connection takes a long time. Keep the ones you use regularly.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Back up models you made yourself\u003C\u002Fstrong>: models created from a Modelfile, models you imported or converted, and weights you fine-tuned may not be downloadable again.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Files are shared\u003C\u002Fstrong>: several Ollama models can share the same data, and recent versions of huggingface_hub also share identical files between repositories. Deleting one model may free less space than its listed size. Do not delete individual files inside \u003Ccode>blobs\u003C\u002Fcode> by hand; other models may still use them.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Ch2 id=\"manual-cleanup\">\u003Ca class=\"heading-anchor\" href=\"#manual-cleanup\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Clean up manually\u003C\u002Fh2>\n\u003Ch3 id=\"ollama\">\u003Ca class=\"heading-anchor\" href=\"#ollama\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Ollama\u003C\u002Fh3>\n\u003Cpre>\u003Ccode class=\"language-bash\"># List downloaded models and their sizes\nollama list\n# Delete a model; include the tag after the colon\nollama rm llama3.2:3b\n# Actual space used by the model folder\ndu -sh ~\u002F.ollama\u002Fmodels\n\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>\u003Ccode>ollama rm\u003C\u002Fcode> deletes only the data unique to that model; anything other models still use is kept. \u003Ccode>ollama ps\u003C\u002Fcode> and \u003Ccode>ollama stop\u003C\u002Fcode> are about memory: the first shows loaded models, the second unloads one, and neither frees disk space. To use a different model folder, see the \u003Ca href=\"https:\u002F\u002Fdocs.ollama.com\u002Ffaq\" target=\"_blank\" rel=\"noopener\">Ollama documentation\u003C\u002Fa>.\u003C\u002Fp>\n\u003Ch3 id=\"hugging-face\">\u003Ca class=\"heading-anchor\" href=\"#hugging-face\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Hugging Face\u003C\u002Fh3>\n\u003Cpre>\u003Ccode class=\"language-bash\"># List cached repositories with size and last access time\nhf cache ls\n# Preview deleting an entire model repository; drop --dry-run to actually delete it\nhf cache rm model\u002Forganization\u002Fmodel-name --dry-run\n# Remove old revisions no branch or tag references, plus leftovers from interrupted downloads\nhf cache prune\n\u003C\u002Fcode>\u003C\u002Fpre>\n\u003Cp>These are the commands in the current huggingface_hub documentation; the \u003Ccode>hf\u003C\u002Fcode> command line is installed with huggingface_hub. Subcommand names have changed between versions, so if a command is not found, run \u003Ccode>hf cache --help\u003C\u002Fcode> or upgrade huggingface_hub. \u003Ccode>~\u002F.cache\u002Fhuggingface\u002Fxet\u003C\u002Fcode> holds the Xet storage cache used for uploads and transfers, and you can delete the whole folder; if an upload was interrupted, it has to start over afterwards.\u003C\u002Fp>\n\u003Ch3 id=\"lm-studio\">\u003Ca class=\"heading-anchor\" href=\"#lm-studio\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>LM Studio\u003C\u002Fh3>\n\u003Cp>LM Studio's \u003Cstrong>My Models\u003C\u002Fstrong> page lists every downloaded model; delete one from its menu. This is also where you change the model folder. In Terminal, \u003Ccode>lms ls\u003C\u002Fcode> lists your local models. If you changed the folder, go by the path shown in My Models.\u003C\u002Fp>\n\u003Ch2 id=\"washbear\">\u003Ca class=\"heading-anchor\" href=\"#washbear\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Review and clean up with WashBear\u003C\u002Fh2>\n\u003Cp>The \u003Cstrong>AI models and caches\u003C\u002Fstrong> category in \u003Ca href=\"\u002Fen\u002Fdocs\u002Fdeep-clean\">Deep Cleanup\u003C\u002Fa> lists models these tools downloaded to their default locations:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Ollama local model library\u003C\u002Fstrong>: Ollama models share files, so WashBear treats the whole library as one item that can only be removed as a whole. If you still use any model in it, leave it unselected and delete models one by one with \u003Ccode>ollama rm\u003C\u002Fcode>. Quit Ollama before cleaning.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Hugging Face local models\u003C\u002Fstrong>: each \u003Ccode>models--\u003C\u002Fcode> repository is listed separately, so you can select only the models you no longer use. Datasets are not included.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>LM Studio local models\u003C\u002Fstrong>: listed per model folder. Only \u003Ccode>~\u002F.lmstudio\u002Fmodels\u003C\u002Fcode> is checked, and LM Studio must be closed before cleaning.\u003C\u002Fli>\n\u003Cli>WashBear also supports models from Whisper, PyTorch Hub, ModelScope, Keras, OpenAI CLIP, TensorFlow Hub, Coqui TTS, GPT4All, Jan, and ComfyUI, plus the \u003Cstrong>Hugging Face download cache\u003C\u002Fstrong> used for transfers.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>\u003Cstrong>Smart recommendation\u003C\u002Fstrong> never preselects these model items; you confirm each one yourself. Two small caches appear under \u003Cstrong>User cache files\u003C\u002Fstrong>: \u003Cstrong>LM Studio cache\u003C\u002Fstrong> and \u003Cstrong>Ollama downloaded updates\u003C\u002Fstrong> (old update archives older than seven days). They contain no models and are preselected by Smart recommendation.\u003C\u002Fp>\n\u003Cdiv class=\"callout callout-note\">\u003Cp class=\"callout-title\">Note\u003C\u002Fp>\u003Cp>WashBear reads environment variables such as \u003Ccode>OLLAMA_MODELS\u003C\u002Fcode>, \u003Ccode>HF_HOME\u003C\u002Fcode>, and \u003Ccode>HF_HUB_CACHE\u003C\u002Fcode> when they are visible to it. Variables set only in your shell configuration, and not with \u003Ccode>launchctl setenv\u003C\u002Fcode>, are not visible to apps opened from Finder or the Dock, so in that case WashBear checks the default locations only. Also, WashBear does not follow symbolic links when measuring Hugging Face repositories, so files that recent caches keep in the shared \u003Ccode>blobs\u003C\u002Fcode> folder are neither counted nor deleted. After removing repositories, run \u003Ccode>hf cache prune\u003C\u002Fcode> to reclaim that space.\u003C\u002Fp>\n\u003C\u002Fdiv>\u003Cp>Cleanup deletes files directly instead of moving them to the Trash, and the result is recorded in \u003Ca href=\"\u002Fen\u002Fdocs\u002Fgetting-started#history\">Operation History\u003C\u002Fa>. To see how much space your models take in total, \u003Ca href=\"\u002Fen\u002Fdownload\">download WashBear\u003C\u002Fa> and run a scan; scanning alone never deletes anything. The demo below uses sample data and is only illustrative.\u003C\u002Fp>\n\u003C!--demo:deepClean-->\u003Ch2 id=\"stray-model-files\">\u003Ca class=\"heading-anchor\" href=\"#stray-model-files\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Find stray model files\u003C\u002Fh2>\n\u003Cp>\u003Ccode>.gguf\u003C\u002Fcode> and \u003Ccode>.safetensors\u003C\u002Fcode> files downloaded straight from the web often end up in Downloads or Documents, outside every tool's model folder. Scan those locations with \u003Ca href=\"\u002Fen\u002Fdocs\u002Flarge-files\">Large File Cleanup\u003C\u002Fa> and filter by \u003Cstrong>AI models\u003C\u002Fstrong> to see large files in formats such as \u003Ccode>.gguf\u003C\u002Fcode>, \u003Ccode>.safetensors\u003C\u002Fcode>, \u003Ccode>.ckpt\u003C\u002Fcode>, \u003Ccode>.pt\u003C\u002Fcode>, and \u003Ccode>.onnx\u003C\u002Fcode> in one place. Ollama's model data has no file extension, so it does not appear in this filter; manage it as described above. To see which folders are largest, use \u003Ca href=\"\u002Fen\u002Fdocs\u002Fdisk-analysis\">Disk Space Analysis\u003C\u002Fa>. The demo below also uses sample data.\u003C\u002Fp>\n\u003C!--demo:largeFiles-->\u003Ch2 id=\"faq\">\u003Ca class=\"heading-anchor\" href=\"#faq\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>FAQ\u003C\u002Fh2>\n\u003Ch3 id=\"faq-reuse\">\u003Ca class=\"heading-anchor\" href=\"#faq-reuse\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Can I use a deleted model again?\u003C\u002Fh3>\n\u003Cp>Yes, just download it again: \u003Ccode>ollama pull\u003C\u002Fcode> for Ollama, the in-app download for LM Studio, and libraries that load models through Hugging Face download them automatically on next use. Models you created or imported yourself are the exception, so back them up first.\u003C\u002Fp>\n\u003Ch3 id=\"faq-sizes-dont-add-up\">\u003Ca class=\"heading-anchor\" href=\"#faq-sizes-dont-add-up\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Why do the sizes in ollama list add up to more than the folder?\u003C\u002Fh3>\n\u003Cp>Models with different names or tags can share the same data, and \u003Ccode>ollama list\u003C\u002Fcode> counts each model separately, so the total is inflated. \u003Ccode>du -sh ~\u002F.ollama\u002Fmodels\u003C\u002Fcode> shows the real usage.\u003C\u002Fp>\n\u003Ch3 id=\"faq-external-drive\">\u003Ca class=\"heading-anchor\" href=\"#faq-external-drive\" aria-hidden=\"true\" tabindex=\"-1\">#\u003C\u002Fa>Can I move models to an external drive?\u003C\u002Fh3>\n\u003Cp>Yes. Set \u003Ccode>OLLAMA_MODELS\u003C\u002Fcode> for Ollama and \u003Ccode>HF_HOME\u003C\u002Fcode> or \u003Ccode>HF_HUB_CACHE\u003C\u002Fcode> for Hugging Face, and change the folder in \u003Cstrong>My Models\u003C\u002Fstrong> for LM Studio. After the move, WashBear finds the new Ollama and Hugging Face locations only if it can read those variables, and custom LM Studio folders are not checked. Use Disk Space Analysis to look at those locations.\u003C\u002Fp>\n",[16,20,23,26,30,33,36,39,42,45,48,51],{"depth":17,"id":18,"text":19},2,"where-models-live","Where models live",{"depth":17,"id":21,"text":22},"before-you-delete","Think before you delete",{"depth":17,"id":24,"text":25},"manual-cleanup","Clean up manually",{"depth":27,"id":28,"text":29},3,"ollama","Ollama",{"depth":27,"id":31,"text":32},"hugging-face","Hugging Face",{"depth":27,"id":34,"text":35},"lm-studio","LM Studio",{"depth":17,"id":37,"text":38},"washbear","Review and clean up with WashBear",{"depth":17,"id":40,"text":41},"stray-model-files","Find stray model files",{"depth":17,"id":43,"text":44},"faq","FAQ",{"depth":27,"id":46,"text":47},"faq-reuse","Can I use a deleted model again?",{"depth":27,"id":49,"text":50},"faq-sizes-dont-add-up","Why do the sizes in ollama list add up to more than the folder?",{"depth":27,"id":52,"text":53},"faq-external-drive","Can I move models to an external drive?","\nA 7B or 8B parameter model still takes 4 to 5 GB after 4-bit quantization. Try a few sizes and quantization levels and tens of gigabytes disappear fast. To make it harder, every tool has its own model folder, and some of them are hidden. This guide lists the default locations of the common tools and how to delete models in each, then explains what WashBear can do for you.\n\n## Where models live {#where-models-live}\n\n| Tool | Default location | Layout |\n| --- | --- | --- |\n| Ollama | `~\u002F.ollama\u002Fmodels` | `blobs` holds model data; `manifests` records which data makes up each model |\n| Hugging Face | `~\u002F.cache\u002Fhuggingface\u002Fhub` | One folder per repository; models start with `models--`, datasets with `datasets--` |\n| LM Studio | `~\u002F.lmstudio\u002Fmodels` | Two levels: publisher, then model |\n\nPython libraries that download models through Hugging Face, such as transformers and diffusers, share the same cache. All of these locations can be changed: Ollama uses the `OLLAMA_MODELS` environment variable, Hugging Face uses `HF_HOME` or `HF_HUB_CACHE`, and LM Studio lets you change it on the app's **My Models** page.\n\n## Think before you delete {#before-you-delete}\n\n- **The cost of downloading again**: models are often several gigabytes, and re-downloading on a slow connection takes a long time. Keep the ones you use regularly.\n- **Back up models you made yourself**: models created from a Modelfile, models you imported or converted, and weights you fine-tuned may not be downloadable again.\n- **Files are shared**: several Ollama models can share the same data, and recent versions of huggingface_hub also share identical files between repositories. Deleting one model may free less space than its listed size. Do not delete individual files inside `blobs` by hand; other models may still use them.\n\n## Clean up manually {#manual-cleanup}\n\n### Ollama {#ollama}\n\n```bash\n# List downloaded models and their sizes\nollama list\n# Delete a model; include the tag after the colon\nollama rm llama3.2:3b\n# Actual space used by the model folder\ndu -sh ~\u002F.ollama\u002Fmodels\n```\n\n`ollama rm` deletes only the data unique to that model; anything other models still use is kept. `ollama ps` and `ollama stop` are about memory: the first shows loaded models, the second unloads one, and neither frees disk space. To use a different model folder, see the [Ollama documentation](https:\u002F\u002Fdocs.ollama.com\u002Ffaq).\n\n### Hugging Face {#hugging-face}\n\n```bash\n# List cached repositories with size and last access time\nhf cache ls\n# Preview deleting an entire model repository; drop --dry-run to actually delete it\nhf cache rm model\u002Forganization\u002Fmodel-name --dry-run\n# Remove old revisions no branch or tag references, plus leftovers from interrupted downloads\nhf cache prune\n```\n\nThese are the commands in the current huggingface_hub documentation; the `hf` command line is installed with huggingface_hub. Subcommand names have changed between versions, so if a command is not found, run `hf cache --help` or upgrade huggingface_hub. `~\u002F.cache\u002Fhuggingface\u002Fxet` holds the Xet storage cache used for uploads and transfers, and you can delete the whole folder; if an upload was interrupted, it has to start over afterwards.\n\n### LM Studio {#lm-studio}\n\nLM Studio's **My Models** page lists every downloaded model; delete one from its menu. This is also where you change the model folder. In Terminal, `lms ls` lists your local models. If you changed the folder, go by the path shown in My Models.\n\n## Review and clean up with WashBear {#washbear}\n\nThe **AI models and caches** category in [Deep Cleanup](\u002Fdocs\u002Fdeep-clean) lists models these tools downloaded to their default locations:\n\n- **Ollama local model library**: Ollama models share files, so WashBear treats the whole library as one item that can only be removed as a whole. If you still use any model in it, leave it unselected and delete models one by one with `ollama rm`. Quit Ollama before cleaning.\n- **Hugging Face local models**: each `models--` repository is listed separately, so you can select only the models you no longer use. Datasets are not included.\n- **LM Studio local models**: listed per model folder. Only `~\u002F.lmstudio\u002Fmodels` is checked, and LM Studio must be closed before cleaning.\n- WashBear also supports models from Whisper, PyTorch Hub, ModelScope, Keras, OpenAI CLIP, TensorFlow Hub, Coqui TTS, GPT4All, Jan, and ComfyUI, plus the **Hugging Face download cache** used for transfers.\n\n**Smart recommendation** never preselects these model items; you confirm each one yourself. Two small caches appear under **User cache files**: **LM Studio cache** and **Ollama downloaded updates** (old update archives older than seven days). They contain no models and are preselected by Smart recommendation.\n\n> [!NOTE]\n> WashBear reads environment variables such as `OLLAMA_MODELS`, `HF_HOME`, and `HF_HUB_CACHE` when they are visible to it. Variables set only in your shell configuration, and not with `launchctl setenv`, are not visible to apps opened from Finder or the Dock, so in that case WashBear checks the default locations only. Also, WashBear does not follow symbolic links when measuring Hugging Face repositories, so files that recent caches keep in the shared `blobs` folder are neither counted nor deleted. After removing repositories, run `hf cache prune` to reclaim that space.\n\nCleanup deletes files directly instead of moving them to the Trash, and the result is recorded in [Operation History](\u002Fdocs\u002Fgetting-started#history). To see how much space your models take in total, [download WashBear](\u002Fdownload) and run a scan; scanning alone never deletes anything. The demo below uses sample data and is only illustrative.\n\n::demo deepClean\n\n## Find stray model files {#stray-model-files}\n\n`.gguf` and `.safetensors` files downloaded straight from the web often end up in Downloads or Documents, outside every tool's model folder. Scan those locations with [Large File Cleanup](\u002Fdocs\u002Flarge-files) and filter by **AI models** to see large files in formats such as `.gguf`, `.safetensors`, `.ckpt`, `.pt`, and `.onnx` in one place. Ollama's model data has no file extension, so it does not appear in this filter; manage it as described above. To see which folders are largest, use [Disk Space Analysis](\u002Fdocs\u002Fdisk-analysis). The demo below also uses sample data.\n\n::demo largeFiles\n\n## FAQ {#faq}\n\n### Can I use a deleted model again? {#faq-reuse}\n\nYes, just download it again: `ollama pull` for Ollama, the in-app download for LM Studio, and libraries that load models through Hugging Face download them automatically on next use. Models you created or imported yourself are the exception, so back them up first.\n\n### Why do the sizes in ollama list add up to more than the folder? {#faq-sizes-dont-add-up}\n\nModels with different names or tags can share the same data, and `ollama list` counts each model separately, so the total is inflated. `du -sh ~\u002F.ollama\u002Fmodels` shows the real usage.\n\n### Can I move models to an external drive? {#faq-external-drive}\n\nYes. Set `OLLAMA_MODELS` for Ollama and `HF_HOME` or `HF_HUB_CACHE` for Hugging Face, and change the folder in **My Models** for LM Studio. After the move, WashBear finds the new Ollama and Hugging Face locations only if it can read those variables, and custom LM Studio folders are not checked. Use Disk Space Analysis to look at those locations.\n",1793258439509]