Understanding How Ollama Stores Models

---

---

AI Summary

Olama Model Storage Summary

• Common Questions:
  • Location of Olama model storage is confusing.
  • Directory contents are unclear without the Olama Manifest.
• Platform Availability:
  • Linux: /user/share/olama/olama/models
  • Mac: /user/root/olama/models
  • Windows support is in progress.
• Manifest Directory:
  • Organized by registry, organization, then model.
  • Currently, a single registry at ol. is available.
  • Official models are under the Library directory.
• Model Directories:
  • Example: llama 2 model directory contains files for each tag pulled.
• Manifest File Structure:
  • Schema version is set to two, related to Docker, not Olama.
  • Media type: application/vnd.docker.distribution.manifest.v1+json
  • Config and layers sections contain media type, digest, and size.
• Layers:
  • First layer usually has the largest size and a unique SHA-256 digest.
  • Digests serve as fingerprints and match file names in the blobs directory.
• File Types:
  • Model files are large and identified by SHA-256 digests.
  • License files have their own media type and digest.
  • Template files also have unique digests, shared across models.
• Model Management:
  • Pulling a new model downloads the manifest and required blobs.
  • Copying a model retains the same model digest.
  • Creating a new model from an existing one uses the same digest.
  • Deleting a model does not remove the blob if it’s still referenced.
  • Blobs are reused across models to save space.
• Conclusion:
  • Models are stored in specific directories based on the platform.
  • Model names are not straightforward without referencing the manifest.
  • Deleting models may not free up space if blobs are shared.
  • The system maintains a single copy of each blob for efficiency.

---