Recover GGUF Model Files: How to Restore Deleted AI Models in LM Studio and Ollama
Recovering GGUF model files from LM Studio or Ollama after an accidental delete or drive failure can save hours of bandwidth and prevent permanent loss of model versions that may no longer be available for download. GGUF files are large, distinctive binary files that data recovery tools can detect reliably — but acting quickly is critical because large files occupy significant disk space that gets reused fast. This guide covers every recovery path for locally stored AI model files.
⚠️ Warning: GGUF files range from 2 GB to over 70 GB each. If you have deleted a large model, stop writing data to that drive immediately — a single large new file or a drive defragmentation can overwrite the exact sectors holding your model. Disconnect the drive if possible and do not install new software to the same partition.
Part 1. Where LM Studio and Ollama Store Model Files
Knowing the default model directories helps you target the right drive and partition during recovery.
LM Studio stores models in:
- Windows:
C:\Users\{username}\.lmstudio\models\ - Mac:
~/.lmstudio/models/ - Each model is a single GGUF file or split GGUF parts named
model-00001-of-00003.gguf
Ollama stores models in:
- Windows:
C:\Users\{username}\.ollama\models\blobs\ - Mac:
~/.ollama/models/blobs/ - Linux:
~/.ollama/models/blobs/ - Ollama stores files as content-addressed blobs (SHA256-named files) rather than human-readable filenames
This distinction matters for recovery: LM Studio files are straightforward GGUF-named files, while Ollama blob files may recover with hash-based names that require matching back to a manifest file to identify the model.
Part 2. GGUF Model Recovery by Application
Table 1: GGUF Model Recovery by Application
| Application | Default Storage Path | File Naming | Recovery Difficulty | Notes |
|---|---|---|---|---|
| LM Studio (Windows) | C:\Users\{user}\.lmstudio\models\ |
{model-name}.gguf |
Low — clear filenames | Easy to identify recovered files |
| LM Studio (Mac) | ~/.lmstudio/models/ |
{model-name}.gguf |
Low — clear filenames | Use Ritridata for Mac |
| Ollama (Windows) | C:\Users\{user}\.ollama\models\blobs\ |
SHA256 hash names | Medium — opaque names | Recover manifest first to identify models |
| Ollama (Mac) | ~/.ollama/models/blobs/ |
SHA256 hash names | Medium | Same as Windows |
| Ollama (Linux) | ~/.ollama/models/blobs/ |
SHA256 hash names | Medium | Linux not directly supported by Ritridata |
| Custom model path (any app) | User-defined | Varies | Low–Medium | Check app settings for configured path |
| Split GGUF files | Any directory | model-XXXXX-of-YYYYY.gguf |
Medium | Must recover all parts for model to load |
| Fine-tuned models (PEFT) | Adjacent to base model | LoRA adapter files | Medium | Recover base GGUF + adapter files together |
💡 Tip: Before running recovery software, check if Ollama has a cached manifest for the deleted model. The manifest file at
~/.ollama/models/manifests/contains the SHA256 hashes of required blobs — recovering this file first helps you identify which blob files belong to which model.
Part 3. GGUF Model File Sizes by Quantization
Understanding model file sizes helps you allocate recovery destination space and identify files during a scan.
Table 2: Model File Sizes by Quantization Level
| Model Size | Quantization | Approximate File Size | Common Use Case |
|---|---|---|---|
| 7B parameters | Q4_K_M | ~4.1 GB | Fast inference, everyday tasks |
| 7B parameters | Q5_K_M | ~5.0 GB | Balanced quality/speed |
| 7B parameters | Q8_0 | ~7.7 GB | High quality, more VRAM needed |
| 13B parameters | Q4_K_M | ~7.9 GB | Better reasoning than 7B |
| 13B parameters | Q8_0 | ~14.7 GB | Near-full precision, 13B |
| 30B parameters | Q4_K_M | ~17.2 GB | Strong general performance |
| 70B parameters | Q4_K_M | ~40.2 GB | Top-tier open model quality |
| 70B parameters | Q8_0 | ~74.5 GB | Near-FP16 quality, very large |
| Mistral 7B Instruct | Q4_K_M | ~4.1 GB | Popular instruction-following model |
| Llama 3 8B | Q4_K_M | ~4.7 GB | Meta's 2024 flagship small model |
When browsing recovery scan results, sorting by file size and looking for files in the 2–75 GB range will quickly surface GGUF model files among other recovered data.
Part 4. Step-by-Step: Recover Deleted GGUF Files with Ritridata
Ritridata recovers large binary files including GGUF models from HDDs, SSDs, and external drives on Windows and Mac.
Step 1: Identify the affected drive Note which drive holds your LM Studio or Ollama models directory. On Windows, this is typically your C: drive or a secondary data drive if you moved models to a larger disk.
Step 2: Stop all disk activity Do not download new models, install software, or save anything to the affected drive. Large GGUF files require many sectors — the more you use the drive, the more sectors get reused.
Step 3: Install Ritridata on a different drive Install on a drive that does not contain the lost models. For most users, this means installing on C: while recovering from a secondary drive, or vice versa.
Step 4: Select the drive and run a deep scan Launch Ritridata and select the affected drive or partition. Run a deep scan — the scan detects large binary file signatures across all sectors.
Step 5: Filter by size and file type Filter results for files above 2 GB. GGUF files will appear among the largest recovered files. Look for filenames containing model names if you lost LM Studio models.
Step 6: Recover to a separate drive with sufficient space GGUF files are large — ensure your recovery destination has at least as much free space as the total size of models you are recovering. Save to a different drive, then copy back to your model directory once confirmed.
💡 Tip: After recovering an Ollama model blob, you may need to recreate the manifest file manually or re-pull the model with
ollama pullto register it with the Ollama registry. The recovered blob file is usable but Ollama needs its manifest to serve the model.
Part 5. Recovering Split GGUF Files
Some models, particularly 70B+ parameter models, are split into multiple GGUF files (e.g., model-00001-of-00003.gguf). All parts must be recovered for the model to load.
During a Ritridata scan, look for all numbered parts of the split model. If one part was overwritten, the model may not load — but you can attempt to re-download only the missing parts if the model is still available on Hugging Face.
💡 Tip: For split GGUF models, sort recovered files alphabetically after filtering by size. All parts of the same model will appear together in sequence, making it easy to confirm all parts were recovered before committing to the restore.
Part 6. Ritridata for AI Model File Recovery
Ritridata is well-suited for GGUF recovery because large binary files are among the most reliably recovered file types — their continuous sector allocation and large size make them easy to detect and reassemble during a deep scan.
The free preview scan shows file sizes and names, letting you confirm you have found the right model files before committing to recovery. This is especially useful when recovering multiple models at once from a drive that held a full local AI setup.
Start a free scan with Ritridata
FAQ
Q1: Can I recover a GGUF file that was deleted from my LM Studio models folder? Yes. GGUF files are large binary files with distinctive headers that recovery tools can detect. As long as the drive has not been heavily used since deletion, Ritridata has a high probability of recovering the complete file.
Q2: How long does it take to scan a 2 TB drive for deleted GGUF files? A deep scan on a 2 TB HDD typically takes 3–6 hours. An SSD of the same size scans faster, typically in 1–3 hours. The scan reads every sector regardless of current file allocation.
Q3: My Ollama models were deleted — will I be able to identify them after recovery?
Ollama stores blobs with SHA256-based filenames. If you also recover the manifest files from ~/.ollama/models/manifests/, you can cross-reference the hashes to identify which blobs belong to which model.
Q4: Can I recover a fine-tuned model (LoRA adapter) along with its base GGUF? Yes, if both were stored on the same drive. LoRA adapter files are typically small (50–500 MB) and recover easily alongside the larger base GGUF file.
Q5: Is there any way to recover GGUF files on Linux? Ritridata currently supports Windows and Mac. For Linux recovery, consider open-source tools like TestDisk or PhotoRec, which support Linux and can recover large binary files including GGUF.
Q6: The model file recovered but LM Studio says it is corrupted — what happened? A partially recovered file can appear as a corruption. This may indicate that some sectors of the file were overwritten before recovery. In this case, re-downloading the model from Hugging Face is likely the fastest path to a working file.
Q7: Can I prevent this by keeping my models on an external drive? Yes. Storing models on a dedicated external drive keeps them separate from your system drive and reduces accidental deletion risk. It also makes recovery easier — you can scan the external drive without touching your system.
Q8: What is the difference between recovering a GGUF from an HDD versus an SSD? HDDs retain deleted data until sectors are reused, which can take days or weeks on a lightly used drive. SSDs with TRIM enabled may purge deleted sectors within hours. Always act faster on SSD — the recovery window is shorter.