New: native Mac app · now on Windows & Linux too

Get back the gigabytes
your AI models are hiding.

Ollama, LM Studio, Hugging Face, ComfyUI and your AI coding tools quietly fill tens of gigabytes in folders you never see. LLM Cleaner finds every model in one scan, shows what is duplicated or forgotten, and frees the space without breaking the tools you still use.

One-time payment · use it on any 2 computers · also runs on NVIDIA DGX Spark · Already bought? Get your installer

Runs 100% on your computerTrash first, restore in one clickNo subscription, no account
LLM Cleaner for Mac: a native window listing 159 GB of local AI models from Ollama, LM Studio, Hugging Face, ComfyUI, Whisper and AI coding caches, with two duplicate Stable Diffusion checkpoints flagged
LLM Cleaner
LLM Cleaner on Windows and Linux: the same scan results with Show, Move and Trash actions on every model
Native on Mac Windows 11 Linux DGX Spark
// Why people buy it

Three things you get back

Disk space is the obvious one. The other two are why people keep the app installed.

20–150 GB

Your disk space

Old quantizations, models you tried once, checkpoints copied into two ComfyUI folders, the same GGUF in LM Studio and Downloads. Most people find tens of gigabytes on the first scan.

Fully reversible

Your confidence

Deleting a random sha256-… blob can break three Ollama models at once. LLM Cleaner knows which layers are shared, sends files to the Trash first, and lets you restore anything in one click.

1 scan

Your afternoon

No more du -sh in eight hidden folders, no reading forum threads about which cache is safe to clear. One list, sorted by size, with the source, path and last-used date for every model.

// How it works

From scan to clean in three steps

One scan. Full picture. Every cleanup action is reversible until you decide otherwise.

01

Scan everything at once

One click searches across Ollama, LM Studio, Hugging Face, Whisper, ComfyUI, AI coding caches, and any custom folder you add. No manual path hunting — it finds models wherever they ended up.

02

See exactly what is worth keeping

Sort by size, filter by tool, spot identical files stored twice (same size plus a content fingerprint, not just the name), and flag 10 GB+ models. Every row shows name, source, path, size, and last modified date.

03

Clean up — restore if anything goes wrong

Move to Trash or the Recycle Bin (not permanent delete), move to an external drive with a link so your tools still work, or export a full Markdown report before touching anything.

// The problem

Local AI tools are powerful. Their storage is invisible.

Ollama, LM Studio, Hugging Face, Whisper, ComfyUI, and AI coding assistants scatter multi-gigabyte model files across hidden system folders. One download, one experiment, one forgotten quantization — and suddenly 40 GB is gone with no easy way to find it.

Models are enormous

A single 7B model takes 4–8 GB. Run a few experiments and your SSD quietly fills up.

Files are scattered everywhere

Seven different apps, seven different hidden cache directories. Nothing tells you the total.

Manual cleanup breaks things

Delete the wrong blob directory and Ollama loses track of models it thought it had.

// The numbers

A single model can take more space than your last three apps combined

These are real file sizes from Hugging Face. One experiment. One forgotten download. One duplicate. That is how 80 GB disappears.

Writing model
Llama 3.3 70B
Q4_K_M · GGUF · current standard
43 GB
Ollama · LM Studio · llama.cpp
Writing model
Mixtral 8×7B
Q4_K_M · GGUF
26.4 GB
LM Studio · llama.cpp
Writing model — extreme
DeepSeek-R1 671B
Q4_K_M · GGUF
404 GB
llama.cpp · Ollama
Vision model
Llama 3.2 Vision 11B
Q4_K_M · GGUF
7.3 GB
Ollama · LM Studio
Image generation model
FLUX.1 dev
safetensors · bf16
23.8 GB
ComfyUI · HuggingFace
Speech recognition model
Whisper large-v3
GGML · ggerganov/whisper.cpp
3.1 GB
Whisper.cpp · Ollama

Downloaded the same GGUF or checkpoint twice, into LM Studio and your Downloads folder? That is a duplicate sitting on your SSD. LLM Cleaner finds it.

// What it does

Every feature exists to free space without breaking anything

LLM Cleaner is not a generic disk analyzer. It understands how each AI tool stores its files, so it can tell you what a file is, who uses it, and what happens if it goes away.

01

One scan across every AI tool

Ollama, LM Studio, Hugging Face, ComfyUI, Whisper, Cursor, Claude Code, Codex, Copilot and loose .gguf / .safetensors / .ckpt files anywhere in your home folder, plus any folder you add.

Why it matters: Respects OLLAMA_MODELS, HF_HOME, HF_HUB_CACHE, LM Studio’s custom download folder and ComfyUI’s extra_model_paths.yaml, so moved setups are found too. On Linux it also reads the system-wide Ollama service.

02

Duplicates found by content, not by name

Files of 50 MB or more that share a size and a content fingerprint are flagged as duplicates, even when they have different names in different tools.

Why it matters: The classic case: the same 7 GB GGUF downloaded once in LM Studio and again into Downloads, or an SDXL checkpoint copied into two ComfyUI installs. That is space you get back with zero loss.

03

Ollama-aware: shared layers are never double-counted

Ollama stores models as anonymous sha256 blobs, and several tags can share the same weights. LLM Cleaner reads the manifests, names every model, and shows only the space you would really get back.

Why it matters: Removing a shared layer by hand breaks every model that uses it. LLM Cleaner keeps shared blobs in place and warns you while Ollama is running.

04

Move models to an external drive, keep them working

Pick models, pick a drive. The files move and a link is left at the old path: a symlink on Mac and Linux, a junction on Windows.

Why it matters: Ollama, LM Studio and ComfyUI keep finding their models with no config change. Your internal SSD stays free for the work you do every day.

05

Trash first, restore with one click

Cleanup goes to the macOS Trash, the Windows Recycle Bin or the Linux Trash, never straight to permanent delete. Trashed models stay visible in the app with a Restore button.

Why it matters: Changed your mind, or a tool complains a week later? Put it back. Permanent deletion always needs a second confirmation.

06

AI coding caches, finally visible

Cursor, Claude Code, Codex, Copilot and Windsurf keep indexes, workspace storage and downloaded models in hidden app folders.

Why it matters: See exactly how big each one is and which are rebuildable caches, so you clear what is safe and leave project memory alone.

07

Export a full inventory report

Save a Markdown report of every model: source, path, size, last modified and duplicate flags.

Why it matters: Useful before a big cleanup, for a team machine, or to decide what to download again later instead of keeping it.

08

Native on every desktop

A SwiftUI app on the Mac, and a light native app on Windows and Linux. Same scanners, same safety rules, same license.

Why it matters: Run it on your MacBook and your Linux GPU box or DGX Spark with one $19 license.

// Why not just use Finder or Explorer?

A file manager shows files. LLM Cleaner knows what they are.

You can clean up by hand, and our guide to finding hidden model files shows how. Here is what you give up.

Finder, Explorer or duGeneric disk analyzersLLM Cleaner
Shows hidden AI folders like ~/.ollama and .cache/huggingfaceWith effort
Names each Ollama blob after the model it belongs to––
Knows which layers are shared between models––
Finds the same model across different tools by content–Some
Groups results by AI tool, with quantization and size––
Moves models to another drive and leaves a working link––
Restore from inside the app after cleanup––
Same app on Mac, Windows and Linux–Rarely
// Who it is for

Built for people who download models faster than they delete them

Developers on a 512 GB MacBook

You pull a new model every week to compare it. LLM Cleaner shows which ones you have not touched in months and moves the keepers to an external SSD.

ComfyUI and Stable Diffusion artists

Checkpoints, LoRAs, VAEs and upscalers pile up across installs. Find the duplicates and the 6 GB checkpoints you tried once.

Windows gamers running local AI

Your C: drive is full and the models are in %USERPROFILE%. Move them to D: in one step, with a junction so nothing breaks. More for Windows →

Researchers on Linux and DGX Spark

Hugging Face caches, vLLM downloads and the system Ollama service, in one view, on ARM64 or x64. More for Linux →

// Coverage

Built for the full local AI stack

LLM Cleaner covers every major tool people actually run locally — including the ones that hide their caches where you'd never think to look.

OllamaOllamaWhere Ollama stores models →LM StudioLM StudioLM Studio disk space guide →Hugging FaceHugging FaceWhere HF models are stored →WhisperWhisperWhere Whisper models live →⬛ComfyUIcheckpoints, LoRAs, VAEsComfyUI models guide →AI Coding CachesCursor · Claude Code · CodexWhy AI agents fill storage →
.ggufGGUF filesquantized LLMs
.stSafetensorsHF & diffusers
~/Custom foldersany path you add
+More comingDocker volumes · NVIDIA NIM
// Models we find

If it runs locally, LLM Cleaner finds it

Every model below can be downloaded and cached on your computer by tools like Ollama, LM Studio, Hugging Face, or ComfyUI. LLM Cleaner detects them all — including ones you downloaded months ago and forgot about.

Writing & Chat
Llama 3.3 70BLlama 3.1 8BMistral 7BMixtral 8×7BQwen 3 8BQwen 3 32BQwen 3 235BDeepSeek-V3Gemma 3 9BGemma 3 27BPhi-4Command R+Falcon 180BYi 34BMiniMax Text
Coding
Qwen 2.5 Coder 32BQwen 2.5 Coder 7BDeepSeek-Coder-V2Mistral Codestral 22BStarCoder2 15BCodeLlama 34BDeepSeek-R1 DistillPhi-4 MiniLlama 3.1 8B Instruct
Reasoning
DeepSeek-R1 671BDeepSeek-R1 Distill 14BDeepSeek-R1 Distill 32BQwen QwQ 32BQwen 3 235B A22BKimi k1.5Marco-o1
Vision & Multimodal
Llama 3.2 Vision 11BLlama 3.2 Vision 90BQwen2.5-VL 7BQwen2.5-VL 72BLLaVA 1.6 34BPixtral 12BMiniMax VLInternVL2 8BMoondream 2PaliGemma
Image Generation
FLUX.1 devFLUX.1 schnellStable Diffusion XLStable Diffusion 3.5SD 3.5 LargeControlNetLoRA adaptersVAE modelsUpscale modelsSDXL Lightning
Speech & Audio
Whisper large-v3Whisper mediumWhisper baseKokoro TTSXTTS-v2MeloTTSBarkWhisperSpeechVosk

Not seeing a model you use? Add its folder with + Folder and LLM Cleaner will scan it on the next run.

// Platforms

One license for every desktop you run AI on

The same scanners and the same safety rules everywhere, built for how each system stores files.

LLM Cleaner for Mac

A native SwiftUI app for macOS 14+. Universal binary for Apple Silicon and Intel, real Trash with Put Back, Show in Finder.

Learn more →

LLM Cleaner for Windows

Windows 10 and 11. Recycle Bin restore, Show in File Explorer, moves models to D: with a junction so tools keep working.

Learn more →

LLM Cleaner for Linux

Debian, Ubuntu, Fedora or any distro via AppImage. Finds the system Ollama service models other cleaners miss.

Learn more →

NVIDIA DGX Spark

The ARM64 build runs on DGX OS. Keep the 4 TB drive for the models you actually use, not three copies of them.

Learn more →
// Download

One license. Every desktop you own.

Buy once for $19 and use LLM Cleaner on any two computers: a Mac, a Windows PC, a Linux box, or an NVIDIA DGX Spark, in any combination.

macOS

macOS 14 Sonoma or later
Apple SiliconIntelmacOS 14+

Native SwiftUI app. One download runs on every Mac.

Already bought? Download
About macOS support →

Windows

Windows 10 or Windows 11
Windows 11Windows 10x64ARM64

Recycle Bin, File Explorer and per-user install. No admin rights needed.

Already bought? Download
About Windows support →

Linux

Ubuntu 22.04+, Debian 12+, Fedora 39+
DebianUbuntuFedorax64ARM64

Debian, Ubuntu, Mint, Pop!_OS, Fedora, or any distro via AppImage.

Already bought? Download
About Linux support →

NVIDIA DGX Spark

DGX OS 7 (Ubuntu 24.04) · any ARM64 Linux
DGX OSARM64JetsonRaspberry Pi 5

Finds models from the system-wide Ollama service and the Hugging Face cache that vLLM and NIM-style workflows fill up.

Already bought? Download
About DGX Spark support →

Version 0.3.0 · Same scanners, duplicate detection and restore on every platform · Runs 100% locally

// Pricing

One-time payment. No subscription.

Pay once, use forever. All current features included and updates through v1.x. No renewal fees, no usage limits.

$19one-time · Mac, Windows & Linux · 2 computers

Price may increase after the early launch period.

By purchasing you agree to our Terms of Use.

  • Mac, Windows and Linux apps: use it on any 2 computers
  • 7 AI tool scanners in one app
  • Duplicate file detection
  • Move models to external drive (link stays)
  • Trash / Recycle Bin with one-click restore
  • AI coding cache detection (Cursor, Claude Code, Codex)
  • Custom folder scanning — persistent across sessions
  • Export full Markdown inventory report
  • Finds models in $OLLAMA_MODELS, $HF_HOME, extra_model_paths.yaml
  • Updates through version 1.x
// Roadmap

What is coming next

NVIDIA NIM & llama.cpp caches

Find model weights cached by NVIDIA NIM containers and llama.cpp downloads, common on DGX Spark and Linux GPU boxes.

Docker model volumes

Detect and manage models stored inside Docker volumes used by local AI tools and self-hosted LLM servers.

Scheduled cleanup suggestions

Automatically surface models you haven't used in 30+ days and give you a one-click review queue.

// Privacy & safety

Designed for trust

LLM Cleaner is a local desktop utility. It has no account system and no telemetry, and it never uploads your models, file names or paths. Read exactly what it sends in the privacy policy.

Runs entirely on your computer — no server involved
No model files uploaded, ever
No automatic deletion of any kind
Every cleanup action requires your confirmation
Move to Trash first — not straight to permanent delete
Restore any trashed model in one click
Reveal exact file paths in Finder or File Explorer before acting
Full Markdown report exportable before any cleanup
// Questions

Frequently asked

No. LLM Cleaner moves files to the Trash first: the macOS Trash, the Windows Recycle Bin, or the Linux Trash you already use. Review them, restore them in one click, or empty the Trash whenever you're ready. Nothing is permanently deleted without your confirmation.

Never. LLM Cleaner scans and cleans entirely on your computer. No model files, file names, paths or scan results ever leave your disk. The app only contacts getllmcleaner.com to activate and periodically re-check your license, and to check for updates. Details are in the privacy policy.

It respects the $OLLAMA_MODELS, $HF_HOME and $HF_HUB_CACHE environment variables automatically, and on Linux it also finds the system-wide Ollama service's models. You can also add any folder manually using the custom folder button — it remembers your paths between scans.

Yes. The Move to Drive feature copies files to your chosen external drive and leaves a link at the original location. Ollama, LM Studio, and other tools keep working without any configuration change.

Ollama · LM Studio · Hugging Face · Whisper · ComfyUI · AI coding caches (Cursor, Claude Code, Codex, Copilot) · loose files anywhere on disk. Detects .gguf, .safetensors, .bin, .pt, .ckpt, .pth, and .mlx files.

macOS 14+ (Apple Silicon and Intel, native SwiftUI app), Windows 10 and 11 (x64 and ARM64), and Linux: Debian, Ubuntu, Mint, Fedora, any distro via AppImage, plus ARM64 machines such as the NVIDIA DGX Spark, Jetson and Raspberry Pi 5. One $19 license works on any two computers, in any combination.

Windows may show a SmartScreen notice for new apps. Click More info, then Run anyway. If macOS says it cannot verify the app, open System Settings, then Privacy & Security, and click Open Anyway. Linux shows no warning.

It depends on how long you have been running local AI. A Mac that has used Ollama and LM Studio for a few months usually has several old models and at least one duplicate, often 20–80 GB. Heavy ComfyUI or Hugging Face users regularly find more than 100 GB. The scan is free to look at in the app before you clean anything.

Yes. One $19 license activates on any two computers, in any combination of Mac, Windows and Linux.

Both. Developers get full path visibility, a Markdown export report, and source-by-source breakdown. Beginners get a safe, guided cleanup flow — nothing disappears unless you explicitly confirm it.

// From the blog

Latest guides on local AI storage

Prefer to clean up by hand? These step-by-step guides show where each tool keeps its models and what is safe to remove.

Nov 1, 2026 · 11 min read

NVIDIA DGX Spark Storage: Where Your Models Go and How to Keep the Drive Clean

Where models land on an NVIDIA DGX Spark: Hugging Face cache, Ollama, NIM, Docker and checkpoints. Commands to measure each, a safe cleanup order and NAS moves.

Nov 1, 2026 · 10 min read

Where Does Ollama Store Models on Linux? (Service vs User Install)

Ollama on Linux stores models in /usr/share/ollama/.ollama/models for the service install, or ~/.ollama/models for your user. How to find, move and delete them.

Nov 1, 2026 · 11 min read

How to Clean the Hugging Face Cache on Linux (vLLM, Transformers, DGX Spark)

Clean the Hugging Face cache on Linux with hf cache ls, rm and prune. Why deleting snapshots frees nothing, HF_HOME vs HF_HUB_CACHE, and moving it to a new disk

Nov 1, 2026 · 10 min read

Where Are AI Models Stored on Windows? Ollama, LM Studio, Hugging Face, ComfyUI

Exact default model folders on Windows 10 and 11 for Ollama, LM Studio, Hugging Face, ComfyUI and Whisper, the env vars that move them, and how to measure them.

Nov 1, 2026 · 10 min read

How to Delete Ollama Models on Windows (and Get the Space Back)

Delete Ollama models on Windows 10 and 11 with ollama rm, find where the blobs live, fix space that does not come back, and uninstall Ollama without leftovers.

Nov 1, 2026 · 10 min read

How Much Disk Space Do You Need for Local LLMs on a Mac?

Real model sizes from Ollama and Hugging Face, mapped to 256 GB, 512 GB, 1 TB and 2 TB Macs, plus how to keep local LLM storage under control as it grows.

Find hidden AI model filesWhere Ollama, LM Studio, Hugging Face and ComfyUI hide models on Mac, Windows and Linux, and how to see them.Free up disk space for local AIA practical plan for keeping a local LLM setup from eating your SSD.Mac "System Data" is huge?How AI models and caches end up counted as System Data, and how to get the space back.C: drive full from AI models?Move Ollama, LM Studio and Hugging Face models to another drive on Windows.

See all guides →

// Ready to clean up?

Stop guessing where
your gigabytes went.

One scan. Every AI model. Every tool. Fully recoverable cleanup.

One-time payment · Use it on 2 computers · No subscription