Skip to content

Model References

A model reference is the user-facing input that tells oLLM what to resolve and run.

Supported forms

Built-in aliases

Examples:

  • llama3-1B-chat
  • llama3-3B-chat
  • llama3-8B-chat
  • gpt-oss-20B
  • qwen3-next-80B
  • gemma3-12B
  • voxtral-small-24B

These aliases are metadata entries, not admission gates for the rest of the runtime.

Hugging Face repo IDs

Examples:

  • Qwen/Qwen2.5-7B-Instruct
  • google/gemma-3-12b-it

These resolve to local materialization paths under --models-dir.

Hugging Face revisions

Examples:

Qwen/Qwen2.5-7B-Instruct@main
Qwen/Qwen2.5-7B-Instruct@refs/pr/12

Revisions are preserved in the materialization path.

Local paths

Examples:

./models/custom-model
~/models/custom-model
/absolute/path/to/model

Resolution outcomes

oLLM resolves a model reference into one of these source kinds:

  • builtin
  • hugging-face
  • local-path
  • opaque

Opaque references are parsed, but not currently resolvable to a runnable runtime or materialization path.