Skip to main content
Import a model to use weights trained outside Oumi in the same workflow as your own models. An imported model can be evaluated against your test sets, compared with models you trained in Oumi, and deployed if its base model supports deployment.

​
WHEN TO IMPORT A MODEL

  • Compare: Evaluate an open-source or third-party fine-tune next to your own models, on the same test set and evaluators.
  • Deploy: Serve a model you trained elsewhere on an Oumi deployment, with its logs and monitoring.

​
IMPORT A MODEL

On the Models page, open the More actions menu and choose a source:
  • Import from HuggingFace: Enter a Display name and the Repo ID, such as meta-llama/Llama-3.2-1B. Oumi imports the repo’s default branch. For a gated or private repo, add a Hugging Face key to your project first.
  • Import from S3: Enter a Display name and a Presigned URL to a weight file or an archive, such as a .zip or .tar.gz. Treat the URL as a secret, because anyone with it can read the file.
Import from HuggingFace dialog with the Display name, Repo ID, and Base model fields The import runs in the background. The model appears on the Models page when it finishes.

​
SET THE BASE MODEL

The base model is the model your weights were trained from. Oumi uses it to format prompts with the right chat template, and to decide whether the model can be deployed. A wrong base model can make evaluations score poorly and can block deployment.
  • Hugging Face imports: Leave Base model on Auto-detect unless your weights derive from a different model.
  • S3 imports: Choose the base model, such as Qwen/Qwen3-0.6B.
  • A base model that isn’t listed: Choose Other… and enter its Hugging Face ID. A model with an unlisted base model may not be deployable.
For a fine-tuned model, the base model is the model you started from, not your checkpoint.

​
LORA ADAPTERS AND FULL WEIGHTS

Oumi detects the kind of weights for you. An import that contains an adapter_config.json file is a LoRA adapter, and anything else is full weights. Whether a model can be deployed depends on both its base model and this kind. A model that can’t be deployed is marked Not deployable in the Deploy Model dialog.

​
IMPORT WITH FAiE

Ask FAiE (Frontier AI Engineer Agent) to import a model, such as “Import meta-llama/Llama-3.2-1B from Hugging Face.” It can continue straight into an evaluation or a deployment.